schema_version: 'v1.2' pipeline: workflow: manifest: capabilities: - workflow_ssa - linear_effects - nested_control_flow - typed_emit - emit_valid_length - loop_induction_values - serving_service_contract - bounded_state_recurrence - parameter_adapters - heterogeneous_adapter_batching inputs: request.input_ids: contract: dtype: int64 rank: 2 shape: - batch - sequence_len batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: prompt_tokens source: kind: request required: true request.max_iterations: contract: dtype: int64 rank: 1 shape: - 1 role: kind: runtime version: '1.0' role: max_output_tokens source: kind: request required: true package.one_token: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: literal required: false default: 1 package.one_step: contract: dtype: int64 rank: 1 shape: - 1 role: kind: opaque source: kind: literal required: false default: 1 package.max_context: contract: dtype: int64 rank: 1 shape: - 1 role: kind: opaque source: kind: literal required: false default: 32768 request.prompt_lengths: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: application name: prompt_lengths required: false default: -1 request.eos_ids: contract: dtype: int64 rank: 2 shape: - batch - num_eos batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: eos_token_ids source: kind: request required: true request.eos_lengths: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: eos_token_lengths source: kind: request required: true request.row_max_iterations: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: application name: row_max_iterations required: false default: -1 request.temperature: contract: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: sampling_temperature source: kind: request required: false default: 1.0 request.top_k: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: sampling_top_k source: kind: request required: false default: 1 request.top_p: contract: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: sampling_top_p source: kind: request required: false default: 1.0 request.min_p: contract: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: sampling_min_p source: kind: request required: false default: 0.0 request.seed: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: seed source: kind: request required: false default: 0 request.rng_counter: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: application name: rng_counter required: false default: 0 package.active: contract: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: literal required: false default: true package.not_done: contract: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: literal required: false default: false package.cache_lengths: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: literal required: false default: 0 package.zero_batch: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: opaque source: kind: literal required: false default: 0 request.adapter_segments: contract: dtype: int64 rank: 2 shape: - batch - 2 batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: adapter_segments source: kind: request request.adapter_counts: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: adapter_counts source: kind: request request.adapter_scales: contract: dtype: float32 rank: 2 shape: - batch - 2 batch_layout: kind: request_aligned axis: 0 role: kind: runtime version: '1.0' role: adapter_scales source: kind: request outputs: tokens: contract: dtype: int64 rank: 2 shape: - batch - generated_sequence batch_layout: kind: request_aligned axis: 0 role: tokens stage: pre_adapter components: model: implementation: kind: onnx artifact: model.onnx ports: roles: input_ids: token_ids attention_mask: attention_mask position_ids: position_ids logits: logits token_sampler: implementation: kind: onnx artifact: policies/token_sampler.onnx ports: inputs: logits: dtype: float32 rank: 2 shape: - batch - vocabulary batch_layout: kind: request_aligned axis: 0 temperature: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 top_k: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 top_p: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 min_p: dtype: float32 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 seed: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 counter: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 done: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: token: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 next_counter: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 contract: id: onnx-genai.token-sampler version: '2' bindings: logits: logits token: token temperature: temperature top_k: top_k top_p: top_p min_p: min_p active: active done: done seed: seed counter: counter next_counter: next_counter parameters: mode: seeded_stochastic batching: per_row inactive_rows: preserve application_overridable: true termination: implementation: kind: onnx artifact: policies/termination.onnx ports: inputs: tokens: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 eos_ids: dtype: int64 rank: 2 shape: - batch - num_eos batch_layout: kind: request_aligned axis: 0 eos_lengths: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 iteration: dtype: int64 rank: 1 shape: - 1 max_iterations: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: done: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 next_active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 continue: dtype: bool rank: 1 shape: - 1 contract: id: onnx-genai.termination-predicate version: '2' bindings: tokens: tokens eos_ids: eos_ids iteration: iteration max_iterations: max_iterations eos_lengths: eos_lengths active: active done: done next_active: next_active continue: continue parameters: batching: per_row inactive_rows: preserve token_state_update: implementation: kind: onnx artifact: policies/token_state_update.onnx ports: inputs: current: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 update: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 done: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: next: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 contract: id: onnx-genai.state-update version: '2' bindings: current: current update: update active: active done: done next: next parameters: batching: per_row inactive_rows: preserve last_token_logits: implementation: kind: onnx artifact: policies/last_token_logits.onnx ports: inputs: logits: dtype: float16 rank: 3 shape: - batch - sequence - vocabulary batch_layout: kind: request_aligned axis: 0 outputs: last_logits: dtype: float32 rank: 2 shape: - batch - vocabulary batch_layout: kind: request_aligned axis: 0 decoder_state_initializer: implementation: kind: onnx artifact: policies/decoder_state_initializer.onnx ports: inputs: prompt_tokens: dtype: int64 rank: 2 shape: - batch - prompt_sequence batch_layout: kind: request_aligned axis: 0 prompt_lengths: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: attention_mask: dtype: int64 rank: 2 shape: - batch - past_seq_len + seq_len batch_layout: kind: request_aligned axis: 0 position_ids: dtype: int64 rank: 2 shape: - batch - prompt_sequence batch_layout: kind: request_aligned axis: 0 body_attention_mask: dtype: int64 rank: 2 shape: - batch - prompt_sequence + 1 batch_layout: kind: request_aligned axis: 0 body_position_ids: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 token_slot: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 generated_lengths: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 past_key_values.0.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.0.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.1.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.1.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.10.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.10.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.11.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.11.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.12.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.12.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.13.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.13.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.14.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.14.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.15.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.15.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.16.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.16.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.17.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.17.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.18.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.18.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.19.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.19.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.2.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.2.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.20.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.20.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.21.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.21.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.22.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.22.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.23.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.23.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.24.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.24.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.25.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.25.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.26.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.26.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.27.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.27.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.3.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.3.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.4.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.4.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.5.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.5.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.6.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.6.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.7.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.7.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.8.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.8.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.9.key: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 past_key_values.9.value: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 decoder_step_update: implementation: kind: onnx artifact: policies/decoder_step_update.onnx ports: inputs: attention_mask: dtype: int64 rank: 2 shape: - batch - context batch_layout: kind: request_aligned axis: 0 position_ids: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 outputs: next_attention_mask: dtype: int64 rank: 2 shape: - batch - context + 1 batch_layout: kind: request_aligned axis: 0 next_position_ids: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 cache_length_update: implementation: kind: onnx artifact: policies/cache_length_update.onnx ports: inputs: left: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 right: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 done: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: total: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 termination_batch_initializer: implementation: kind: onnx artifact: policies/termination_batch_initializer.onnx ports: inputs: input_eos_ids: dtype: int64 rank: 2 shape: - batch - num_eos batch_layout: kind: request_aligned axis: 0 input_eos_lengths: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 input_max_iterations: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 fallback_max_iterations: dtype: int64 rank: 1 shape: - 1 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: row_eos_ids: dtype: int64 rank: 2 shape: - batch - num_eos batch_layout: kind: request_aligned axis: 0 eos_lengths: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 max_iterations: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 token_to_slot: implementation: kind: onnx artifact: policies/token_to_slot.onnx ports: inputs: token: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: slot: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 generated_length_update: implementation: kind: onnx artifact: policies/generated_length_update.onnx ports: inputs: left: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 right: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 active: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 done: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 outputs: total: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 state: token: contract: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: initializer.token_slot recurrence: kind: invariant logits: contract: dtype: float32 rank: 2 shape: - batch - 151936 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.last_logits recurrence: kind: invariant generated_lengths: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 class: semantic scope: invocation initializer: initializer.generated_lengths recurrence: kind: invariant active: contract: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 class: semantic scope: invocation initializer: package.active recurrence: kind: invariant done: contract: dtype: bool rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 class: semantic scope: invocation initializer: package.not_done recurrence: kind: invariant accepted_len: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 class: semantic scope: invocation initializer: package.zero_batch recurrence: kind: invariant cache_lengths: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 class: semantic scope: invocation initializer: package.cache_lengths recurrence: kind: invariant rng_counter: contract: dtype: int64 rank: 1 shape: - batch batch_layout: kind: request_aligned axis: 0 scope: invocation class: semantic initializer: request.rng_counter recurrence: kind: invariant attention_mask: contract: dtype: int64 rank: 2 shape: - batch - context batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: initializer.body_attention_mask recurrence: kind: growing axis: 1 increment: package.one_step max: package.max_context position_ids: contract: dtype: int64 rank: 2 shape: - batch - 1 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: initializer.body_position_ids recurrence: kind: invariant cache_0: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.0.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_1: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.0.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_2: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.1.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_3: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.1.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_4: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.2.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_5: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.2.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_6: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.3.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_7: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.3.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_8: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.4.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_9: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.4.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_10: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.5.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_11: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.5.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_12: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.6.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_13: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.6.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_14: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.7.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_15: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.7.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_16: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.8.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_17: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.8.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_18: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.9.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_19: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.9.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_20: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.10.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_21: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.10.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_22: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.11.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_23: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.11.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_24: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.12.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_25: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.12.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_26: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.13.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_27: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.13.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_28: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.14.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_29: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.14.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_30: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.15.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_31: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.15.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_32: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.16.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_33: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.16.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_34: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.17.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_35: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.17.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_36: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.18.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_37: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.18.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_38: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.19.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_39: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.19.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_40: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.20.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_41: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.20.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_42: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.21.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_43: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.21.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_44: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.22.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_45: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.22.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_46: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.23.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_47: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.23.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_48: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.24.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_49: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.24.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_50: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.25.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_51: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.25.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_52: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.26.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_53: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.26.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_54: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.27.key recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache cache_55: contract: dtype: float16 rank: 4 shape: - batch - 2 - past_sequence_len - 128 batch_layout: kind: request_aligned axis: 0 scope: invocation initializer: decoder.setup.present.27.value recurrence: kind: bounded axis: 2 max: package.max_context management: runtime release_boundary: invocation service_group: decoder_cache serving: active: active done: done accepted_len: accepted_len state_service: groups: decoder_cache: kind: full_attention sequence_axis: 2 layout: bnsh aliasing: forbidden reuse: prefix_reusable: true evictable_prefix: false ports: model: cache_0: input: past_key_values.0.key output: present.0.key role: key layer: 0 cache_1: input: past_key_values.0.value output: present.0.value role: value layer: 0 cache_2: input: past_key_values.1.key output: present.1.key role: key layer: 1 cache_3: input: past_key_values.1.value output: present.1.value role: value layer: 1 cache_4: input: past_key_values.2.key output: present.2.key role: key layer: 2 cache_5: input: past_key_values.2.value output: present.2.value role: value layer: 2 cache_6: input: past_key_values.3.key output: present.3.key role: key layer: 3 cache_7: input: past_key_values.3.value output: present.3.value role: value layer: 3 cache_8: input: past_key_values.4.key output: present.4.key role: key layer: 4 cache_9: input: past_key_values.4.value output: present.4.value role: value layer: 4 cache_10: input: past_key_values.5.key output: present.5.key role: key layer: 5 cache_11: input: past_key_values.5.value output: present.5.value role: value layer: 5 cache_12: input: past_key_values.6.key output: present.6.key role: key layer: 6 cache_13: input: past_key_values.6.value output: present.6.value role: value layer: 6 cache_14: input: past_key_values.7.key output: present.7.key role: key layer: 7 cache_15: input: past_key_values.7.value output: present.7.value role: value layer: 7 cache_16: input: past_key_values.8.key output: present.8.key role: key layer: 8 cache_17: input: past_key_values.8.value output: present.8.value role: value layer: 8 cache_18: input: past_key_values.9.key output: present.9.key role: key layer: 9 cache_19: input: past_key_values.9.value output: present.9.value role: value layer: 9 cache_20: input: past_key_values.10.key output: present.10.key role: key layer: 10 cache_21: input: past_key_values.10.value output: present.10.value role: value layer: 10 cache_22: input: past_key_values.11.key output: present.11.key role: key layer: 11 cache_23: input: past_key_values.11.value output: present.11.value role: value layer: 11 cache_24: input: past_key_values.12.key output: present.12.key role: key layer: 12 cache_25: input: past_key_values.12.value output: present.12.value role: value layer: 12 cache_26: input: past_key_values.13.key output: present.13.key role: key layer: 13 cache_27: input: past_key_values.13.value output: present.13.value role: value layer: 13 cache_28: input: past_key_values.14.key output: present.14.key role: key layer: 14 cache_29: input: past_key_values.14.value output: present.14.value role: value layer: 14 cache_30: input: past_key_values.15.key output: present.15.key role: key layer: 15 cache_31: input: past_key_values.15.value output: present.15.value role: value layer: 15 cache_32: input: past_key_values.16.key output: present.16.key role: key layer: 16 cache_33: input: past_key_values.16.value output: present.16.value role: value layer: 16 cache_34: input: past_key_values.17.key output: present.17.key role: key layer: 17 cache_35: input: past_key_values.17.value output: present.17.value role: value layer: 17 cache_36: input: past_key_values.18.key output: present.18.key role: key layer: 18 cache_37: input: past_key_values.18.value output: present.18.value role: value layer: 18 cache_38: input: past_key_values.19.key output: present.19.key role: key layer: 19 cache_39: input: past_key_values.19.value output: present.19.value role: value layer: 19 cache_40: input: past_key_values.20.key output: present.20.key role: key layer: 20 cache_41: input: past_key_values.20.value output: present.20.value role: value layer: 20 cache_42: input: past_key_values.21.key output: present.21.key role: key layer: 21 cache_43: input: past_key_values.21.value output: present.21.value role: value layer: 21 cache_44: input: past_key_values.22.key output: present.22.key role: key layer: 22 cache_45: input: past_key_values.22.value output: present.22.value role: value layer: 22 cache_46: input: past_key_values.23.key output: present.23.key role: key layer: 23 cache_47: input: past_key_values.23.value output: present.23.value role: value layer: 23 cache_48: input: past_key_values.24.key output: present.24.key role: key layer: 24 cache_49: input: past_key_values.24.value output: present.24.value role: value layer: 24 cache_50: input: past_key_values.25.key output: present.25.key role: key layer: 25 cache_51: input: past_key_values.25.value output: present.25.value role: value layer: 25 cache_52: input: past_key_values.26.key output: present.26.key role: key layer: 26 cache_53: input: past_key_values.26.value output: present.26.value role: value layer: 26 cache_54: input: past_key_values.27.key output: present.27.key role: key layer: 27 cache_55: input: past_key_values.27.value output: present.27.value role: value layer: 27 steps: - kind: loop setup: - kind: invoke component: decoder_state_initializer inputs: prompt_tokens: request.input_ids prompt_lengths: request.prompt_lengths outputs: attention_mask: initializer.attention_mask body_attention_mask: initializer.body_attention_mask token_slot: initializer.token_slot generated_lengths: initializer.generated_lengths position_ids: initializer.position_ids body_position_ids: initializer.body_position_ids past_key_values.0.key: initializer.past_key_values.0.key past_key_values.0.value: initializer.past_key_values.0.value past_key_values.1.key: initializer.past_key_values.1.key past_key_values.1.value: initializer.past_key_values.1.value past_key_values.10.key: initializer.past_key_values.10.key past_key_values.10.value: initializer.past_key_values.10.value past_key_values.11.key: initializer.past_key_values.11.key past_key_values.11.value: initializer.past_key_values.11.value past_key_values.12.key: initializer.past_key_values.12.key past_key_values.12.value: initializer.past_key_values.12.value past_key_values.13.key: initializer.past_key_values.13.key past_key_values.13.value: initializer.past_key_values.13.value past_key_values.14.key: initializer.past_key_values.14.key past_key_values.14.value: initializer.past_key_values.14.value past_key_values.15.key: initializer.past_key_values.15.key past_key_values.15.value: initializer.past_key_values.15.value past_key_values.16.key: initializer.past_key_values.16.key past_key_values.16.value: initializer.past_key_values.16.value past_key_values.17.key: initializer.past_key_values.17.key past_key_values.17.value: initializer.past_key_values.17.value past_key_values.18.key: initializer.past_key_values.18.key past_key_values.18.value: initializer.past_key_values.18.value past_key_values.19.key: initializer.past_key_values.19.key past_key_values.19.value: initializer.past_key_values.19.value past_key_values.2.key: initializer.past_key_values.2.key past_key_values.2.value: initializer.past_key_values.2.value past_key_values.20.key: initializer.past_key_values.20.key past_key_values.20.value: initializer.past_key_values.20.value past_key_values.21.key: initializer.past_key_values.21.key past_key_values.21.value: initializer.past_key_values.21.value past_key_values.22.key: initializer.past_key_values.22.key past_key_values.22.value: initializer.past_key_values.22.value past_key_values.23.key: initializer.past_key_values.23.key past_key_values.23.value: initializer.past_key_values.23.value past_key_values.24.key: initializer.past_key_values.24.key past_key_values.24.value: initializer.past_key_values.24.value past_key_values.25.key: initializer.past_key_values.25.key past_key_values.25.value: initializer.past_key_values.25.value past_key_values.26.key: initializer.past_key_values.26.key past_key_values.26.value: initializer.past_key_values.26.value past_key_values.27.key: initializer.past_key_values.27.key past_key_values.27.value: initializer.past_key_values.27.value past_key_values.3.key: initializer.past_key_values.3.key past_key_values.3.value: initializer.past_key_values.3.value past_key_values.4.key: initializer.past_key_values.4.key past_key_values.4.value: initializer.past_key_values.4.value past_key_values.5.key: initializer.past_key_values.5.key past_key_values.5.value: initializer.past_key_values.5.value past_key_values.6.key: initializer.past_key_values.6.key past_key_values.6.value: initializer.past_key_values.6.value past_key_values.7.key: initializer.past_key_values.7.key past_key_values.7.value: initializer.past_key_values.7.value past_key_values.8.key: initializer.past_key_values.8.key past_key_values.8.value: initializer.past_key_values.8.value past_key_values.9.key: initializer.past_key_values.9.key past_key_values.9.value: initializer.past_key_values.9.value - kind: invoke component: model inputs: input_ids: request.input_ids past_key_values.0.key: initializer.past_key_values.0.key past_key_values.0.value: initializer.past_key_values.0.value past_key_values.1.key: initializer.past_key_values.1.key past_key_values.1.value: initializer.past_key_values.1.value past_key_values.2.key: initializer.past_key_values.2.key past_key_values.2.value: initializer.past_key_values.2.value past_key_values.3.key: initializer.past_key_values.3.key past_key_values.3.value: initializer.past_key_values.3.value past_key_values.4.key: initializer.past_key_values.4.key past_key_values.4.value: initializer.past_key_values.4.value past_key_values.5.key: initializer.past_key_values.5.key past_key_values.5.value: initializer.past_key_values.5.value past_key_values.6.key: initializer.past_key_values.6.key past_key_values.6.value: initializer.past_key_values.6.value past_key_values.7.key: initializer.past_key_values.7.key past_key_values.7.value: initializer.past_key_values.7.value past_key_values.8.key: initializer.past_key_values.8.key past_key_values.8.value: initializer.past_key_values.8.value past_key_values.9.key: initializer.past_key_values.9.key past_key_values.9.value: initializer.past_key_values.9.value past_key_values.10.key: initializer.past_key_values.10.key past_key_values.10.value: initializer.past_key_values.10.value past_key_values.11.key: initializer.past_key_values.11.key past_key_values.11.value: initializer.past_key_values.11.value past_key_values.12.key: initializer.past_key_values.12.key past_key_values.12.value: initializer.past_key_values.12.value past_key_values.13.key: initializer.past_key_values.13.key past_key_values.13.value: initializer.past_key_values.13.value past_key_values.14.key: initializer.past_key_values.14.key past_key_values.14.value: initializer.past_key_values.14.value past_key_values.15.key: initializer.past_key_values.15.key past_key_values.15.value: initializer.past_key_values.15.value past_key_values.16.key: initializer.past_key_values.16.key past_key_values.16.value: initializer.past_key_values.16.value past_key_values.17.key: initializer.past_key_values.17.key past_key_values.17.value: initializer.past_key_values.17.value past_key_values.18.key: initializer.past_key_values.18.key past_key_values.18.value: initializer.past_key_values.18.value past_key_values.19.key: initializer.past_key_values.19.key past_key_values.19.value: initializer.past_key_values.19.value past_key_values.20.key: initializer.past_key_values.20.key past_key_values.20.value: initializer.past_key_values.20.value past_key_values.21.key: initializer.past_key_values.21.key past_key_values.21.value: initializer.past_key_values.21.value past_key_values.22.key: initializer.past_key_values.22.key past_key_values.22.value: initializer.past_key_values.22.value past_key_values.23.key: initializer.past_key_values.23.key past_key_values.23.value: initializer.past_key_values.23.value past_key_values.24.key: initializer.past_key_values.24.key past_key_values.24.value: initializer.past_key_values.24.value past_key_values.25.key: initializer.past_key_values.25.key past_key_values.25.value: initializer.past_key_values.25.value past_key_values.26.key: initializer.past_key_values.26.key past_key_values.26.value: initializer.past_key_values.26.value past_key_values.27.key: initializer.past_key_values.27.key past_key_values.27.value: initializer.past_key_values.27.value attention_mask: initializer.attention_mask position_ids: initializer.position_ids outputs: logits: decoder.setup.logits present.0.key: decoder.setup.present.0.key present.0.value: decoder.setup.present.0.value present.1.key: decoder.setup.present.1.key present.1.value: decoder.setup.present.1.value present.2.key: decoder.setup.present.2.key present.2.value: decoder.setup.present.2.value present.3.key: decoder.setup.present.3.key present.3.value: decoder.setup.present.3.value present.4.key: decoder.setup.present.4.key present.4.value: decoder.setup.present.4.value present.5.key: decoder.setup.present.5.key present.5.value: decoder.setup.present.5.value present.6.key: decoder.setup.present.6.key present.6.value: decoder.setup.present.6.value present.7.key: decoder.setup.present.7.key present.7.value: decoder.setup.present.7.value present.8.key: decoder.setup.present.8.key present.8.value: decoder.setup.present.8.value present.9.key: decoder.setup.present.9.key present.9.value: decoder.setup.present.9.value present.10.key: decoder.setup.present.10.key present.10.value: decoder.setup.present.10.value present.11.key: decoder.setup.present.11.key present.11.value: decoder.setup.present.11.value present.12.key: decoder.setup.present.12.key present.12.value: decoder.setup.present.12.value present.13.key: decoder.setup.present.13.key present.13.value: decoder.setup.present.13.value present.14.key: decoder.setup.present.14.key present.14.value: decoder.setup.present.14.value present.15.key: decoder.setup.present.15.key present.15.value: decoder.setup.present.15.value present.16.key: decoder.setup.present.16.key present.16.value: decoder.setup.present.16.value present.17.key: decoder.setup.present.17.key present.17.value: decoder.setup.present.17.value present.18.key: decoder.setup.present.18.key present.18.value: decoder.setup.present.18.value present.19.key: decoder.setup.present.19.key present.19.value: decoder.setup.present.19.value present.20.key: decoder.setup.present.20.key present.20.value: decoder.setup.present.20.value present.21.key: decoder.setup.present.21.key present.21.value: decoder.setup.present.21.value present.22.key: decoder.setup.present.22.key present.22.value: decoder.setup.present.22.value present.23.key: decoder.setup.present.23.key present.23.value: decoder.setup.present.23.value present.24.key: decoder.setup.present.24.key present.24.value: decoder.setup.present.24.value present.25.key: decoder.setup.present.25.key present.25.value: decoder.setup.present.25.value present.26.key: decoder.setup.present.26.key present.26.value: decoder.setup.present.26.value present.27.key: decoder.setup.present.27.key present.27.value: decoder.setup.present.27.value - kind: invoke component: termination_batch_initializer inputs: input_eos_ids: request.eos_ids input_eos_lengths: request.eos_lengths input_max_iterations: request.row_max_iterations fallback_max_iterations: request.max_iterations active: package.active outputs: row_eos_ids: termination.eos_ids eos_lengths: termination.eos_lengths max_iterations: termination.max_iterations - kind: invoke component: last_token_logits inputs: logits: decoder.setup.logits outputs: last_logits: decoder.setup.last_logits steps: - kind: invoke component: token_sampler inputs: logits: logits temperature: request.temperature top_k: request.top_k top_p: request.top_p min_p: request.min_p seed: request.seed counter: rng_counter active: active done: done outputs: token: sample.body next_counter: sample.next_counter - kind: invoke component: token_to_slot inputs: token: sample.body outputs: slot: sample.slot - kind: invoke component: generated_length_update inputs: left: generated_lengths right: package.one_token active: active done: done outputs: total: token.next_lengths - kind: invoke component: generated_length_update inputs: left: package.zero_batch right: package.one_token active: active done: done outputs: total: token.emitted_length - kind: invoke component: token_state_update inputs: current: token update: sample.slot active: active done: done outputs: next: token.body - kind: invoke component: termination inputs: tokens: sample.body eos_ids: termination.eos_ids eos_lengths: termination.eos_lengths iteration: loop.iteration max_iterations: termination.max_iterations active: active outputs: done: loop.done continue: loop.continue next_active: loop.next_active - kind: invoke component: cache_length_update inputs: left: cache_lengths right: package.one_token active: active done: done outputs: total: cache_lengths.next - kind: invoke component: cache_length_update inputs: left: package.zero_batch right: package.one_token active: active done: done outputs: total: accepted_len.next - kind: emit value: token.body output: tokens mode: append valid_length: token.emitted_length when: active - kind: invoke component: model inputs: input_ids: token.body past_key_values.0.key: cache_0 past_key_values.0.value: cache_1 past_key_values.1.key: cache_2 past_key_values.1.value: cache_3 past_key_values.2.key: cache_4 past_key_values.2.value: cache_5 past_key_values.3.key: cache_6 past_key_values.3.value: cache_7 past_key_values.4.key: cache_8 past_key_values.4.value: cache_9 past_key_values.5.key: cache_10 past_key_values.5.value: cache_11 past_key_values.6.key: cache_12 past_key_values.6.value: cache_13 past_key_values.7.key: cache_14 past_key_values.7.value: cache_15 past_key_values.8.key: cache_16 past_key_values.8.value: cache_17 past_key_values.9.key: cache_18 past_key_values.9.value: cache_19 past_key_values.10.key: cache_20 past_key_values.10.value: cache_21 past_key_values.11.key: cache_22 past_key_values.11.value: cache_23 past_key_values.12.key: cache_24 past_key_values.12.value: cache_25 past_key_values.13.key: cache_26 past_key_values.13.value: cache_27 past_key_values.14.key: cache_28 past_key_values.14.value: cache_29 past_key_values.15.key: cache_30 past_key_values.15.value: cache_31 past_key_values.16.key: cache_32 past_key_values.16.value: cache_33 past_key_values.17.key: cache_34 past_key_values.17.value: cache_35 past_key_values.18.key: cache_36 past_key_values.18.value: cache_37 past_key_values.19.key: cache_38 past_key_values.19.value: cache_39 past_key_values.20.key: cache_40 past_key_values.20.value: cache_41 past_key_values.21.key: cache_42 past_key_values.21.value: cache_43 past_key_values.22.key: cache_44 past_key_values.22.value: cache_45 past_key_values.23.key: cache_46 past_key_values.23.value: cache_47 past_key_values.24.key: cache_48 past_key_values.24.value: cache_49 past_key_values.25.key: cache_50 past_key_values.25.value: cache_51 past_key_values.26.key: cache_52 past_key_values.26.value: cache_53 past_key_values.27.key: cache_54 past_key_values.27.value: cache_55 attention_mask: attention_mask position_ids: position_ids outputs: logits: decoder.body.logits present.0.key: decoder.body.present.0.key present.0.value: decoder.body.present.0.value present.1.key: decoder.body.present.1.key present.1.value: decoder.body.present.1.value present.2.key: decoder.body.present.2.key present.2.value: decoder.body.present.2.value present.3.key: decoder.body.present.3.key present.3.value: decoder.body.present.3.value present.4.key: decoder.body.present.4.key present.4.value: decoder.body.present.4.value present.5.key: decoder.body.present.5.key present.5.value: decoder.body.present.5.value present.6.key: decoder.body.present.6.key present.6.value: decoder.body.present.6.value present.7.key: decoder.body.present.7.key present.7.value: decoder.body.present.7.value present.8.key: decoder.body.present.8.key present.8.value: decoder.body.present.8.value present.9.key: decoder.body.present.9.key present.9.value: decoder.body.present.9.value present.10.key: decoder.body.present.10.key present.10.value: decoder.body.present.10.value present.11.key: decoder.body.present.11.key present.11.value: decoder.body.present.11.value present.12.key: decoder.body.present.12.key present.12.value: decoder.body.present.12.value present.13.key: decoder.body.present.13.key present.13.value: decoder.body.present.13.value present.14.key: decoder.body.present.14.key present.14.value: decoder.body.present.14.value present.15.key: decoder.body.present.15.key present.15.value: decoder.body.present.15.value present.16.key: decoder.body.present.16.key present.16.value: decoder.body.present.16.value present.17.key: decoder.body.present.17.key present.17.value: decoder.body.present.17.value present.18.key: decoder.body.present.18.key present.18.value: decoder.body.present.18.value present.19.key: decoder.body.present.19.key present.19.value: decoder.body.present.19.value present.20.key: decoder.body.present.20.key present.20.value: decoder.body.present.20.value present.21.key: decoder.body.present.21.key present.21.value: decoder.body.present.21.value present.22.key: decoder.body.present.22.key present.22.value: decoder.body.present.22.value present.23.key: decoder.body.present.23.key present.23.value: decoder.body.present.23.value present.24.key: decoder.body.present.24.key present.24.value: decoder.body.present.24.value present.25.key: decoder.body.present.25.key present.25.value: decoder.body.present.25.value present.26.key: decoder.body.present.26.key present.26.value: decoder.body.present.26.value present.27.key: decoder.body.present.27.key present.27.value: decoder.body.present.27.value - kind: invoke component: last_token_logits inputs: logits: decoder.body.logits outputs: last_logits: decoder.body.last_logits - kind: invoke component: decoder_step_update inputs: attention_mask: attention_mask position_ids: position_ids outputs: next_attention_mask: decoder_step.body_attention_mask next_position_ids: decoder_step.body_position_ids continue_when: active max_iterations: request.max_iterations carried: - cell: token next: token.body - cell: logits next: decoder.body.last_logits - cell: generated_lengths next: token.next_lengths - cell: active next: loop.next_active - cell: done next: loop.done - cell: cache_lengths next: cache_lengths.next - cell: accepted_len next: accepted_len.next - cell: rng_counter next: sample.next_counter - cell: attention_mask next: decoder_step.body_attention_mask - cell: position_ids next: decoder_step.body_position_ids - cell: cache_0 next: decoder.body.present.0.key - cell: cache_1 next: decoder.body.present.0.value - cell: cache_2 next: decoder.body.present.1.key - cell: cache_3 next: decoder.body.present.1.value - cell: cache_4 next: decoder.body.present.2.key - cell: cache_5 next: decoder.body.present.2.value - cell: cache_6 next: decoder.body.present.3.key - cell: cache_7 next: decoder.body.present.3.value - cell: cache_8 next: decoder.body.present.4.key - cell: cache_9 next: decoder.body.present.4.value - cell: cache_10 next: decoder.body.present.5.key - cell: cache_11 next: decoder.body.present.5.value - cell: cache_12 next: decoder.body.present.6.key - cell: cache_13 next: decoder.body.present.6.value - cell: cache_14 next: decoder.body.present.7.key - cell: cache_15 next: decoder.body.present.7.value - cell: cache_16 next: decoder.body.present.8.key - cell: cache_17 next: decoder.body.present.8.value - cell: cache_18 next: decoder.body.present.9.key - cell: cache_19 next: decoder.body.present.9.value - cell: cache_20 next: decoder.body.present.10.key - cell: cache_21 next: decoder.body.present.10.value - cell: cache_22 next: decoder.body.present.11.key - cell: cache_23 next: decoder.body.present.11.value - cell: cache_24 next: decoder.body.present.12.key - cell: cache_25 next: decoder.body.present.12.value - cell: cache_26 next: decoder.body.present.13.key - cell: cache_27 next: decoder.body.present.13.value - cell: cache_28 next: decoder.body.present.14.key - cell: cache_29 next: decoder.body.present.14.value - cell: cache_30 next: decoder.body.present.15.key - cell: cache_31 next: decoder.body.present.15.value - cell: cache_32 next: decoder.body.present.16.key - cell: cache_33 next: decoder.body.present.16.value - cell: cache_34 next: decoder.body.present.17.key - cell: cache_35 next: decoder.body.present.17.value - cell: cache_36 next: decoder.body.present.18.key - cell: cache_37 next: decoder.body.present.18.value - cell: cache_38 next: decoder.body.present.19.key - cell: cache_39 next: decoder.body.present.19.value - cell: cache_40 next: decoder.body.present.20.key - cell: cache_41 next: decoder.body.present.20.value - cell: cache_42 next: decoder.body.present.21.key - cell: cache_43 next: decoder.body.present.21.value - cell: cache_44 next: decoder.body.present.22.key - cell: cache_45 next: decoder.body.present.22.value - cell: cache_46 next: decoder.body.present.23.key - cell: cache_47 next: decoder.body.present.23.value - cell: cache_48 next: decoder.body.present.24.key - cell: cache_49 next: decoder.body.present.24.value - cell: cache_50 next: decoder.body.present.25.key - cell: cache_51 next: decoder.body.present.25.value - cell: cache_52 next: decoder.body.present.26.key - cell: cache_53 next: decoder.body.present.26.value - cell: cache_54 next: decoder.body.present.27.key - cell: cache_55 next: decoder.body.present.27.value termination: generation_eos iteration: value: loop.iteration contract: dtype: int64 rank: 1 shape: - 1 adapters: target_manifest: targets: - id: model.layers.0.mlp.down_proj component: model initializer: model.layers.0.mlp.down_proj.weight node_name: model/layers.0/mlp/down_proj/MatMul_node_37/adapter_weight_transpose output_name: model.layers.0.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.mlp.gate_proj component: model initializer: model.layers.0.mlp.gate_proj.weight node_name: model/layers.0/mlp/gate_proj/MatMul_node_31/adapter_weight_transpose output_name: model.layers.0.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.mlp.up_proj component: model initializer: model.layers.0.mlp.up_proj.weight node_name: model/layers.0/mlp/up_proj/MatMul_node_34/adapter_weight_transpose output_name: model.layers.0.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.self_attn.k_proj component: model initializer: model.layers.0.self_attn.k_proj.weight node_name: model/layers.0/self_attn/k_proj/MatMul_node_18/adapter_weight_transpose output_name: model.layers.0.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.self_attn.o_proj component: model initializer: model.layers.0.self_attn.o_proj.weight node_name: model/layers.0/self_attn/o_proj/MatMul_node_27/adapter_weight_transpose output_name: model.layers.0.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.self_attn.q_proj component: model initializer: model.layers.0.self_attn.q_proj.weight node_name: model/layers.0/self_attn/q_proj/MatMul_node_15/adapter_weight_transpose output_name: model.layers.0.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.0.self_attn.v_proj component: model initializer: model.layers.0.self_attn.v_proj.weight node_name: model/layers.0/self_attn/v_proj/MatMul_node_21/adapter_weight_transpose output_name: model.layers.0.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 0 rank: 16 alpha: 32.0 - id: model.layers.1.mlp.down_proj component: model initializer: model.layers.1.mlp.down_proj.weight node_name: model/layers.1/mlp/down_proj/MatMul_node_63/adapter_weight_transpose output_name: model.layers.1.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.mlp.gate_proj component: model initializer: model.layers.1.mlp.gate_proj.weight node_name: model/layers.1/mlp/gate_proj/MatMul_node_57/adapter_weight_transpose output_name: model.layers.1.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.mlp.up_proj component: model initializer: model.layers.1.mlp.up_proj.weight node_name: model/layers.1/mlp/up_proj/MatMul_node_60/adapter_weight_transpose output_name: model.layers.1.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.self_attn.k_proj component: model initializer: model.layers.1.self_attn.k_proj.weight node_name: model/layers.1/self_attn/k_proj/MatMul_node_44/adapter_weight_transpose output_name: model.layers.1.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.self_attn.o_proj component: model initializer: model.layers.1.self_attn.o_proj.weight node_name: model/layers.1/self_attn/o_proj/MatMul_node_53/adapter_weight_transpose output_name: model.layers.1.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.self_attn.q_proj component: model initializer: model.layers.1.self_attn.q_proj.weight node_name: model/layers.1/self_attn/q_proj/MatMul_node_41/adapter_weight_transpose output_name: model.layers.1.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.1.self_attn.v_proj component: model initializer: model.layers.1.self_attn.v_proj.weight node_name: model/layers.1/self_attn/v_proj/MatMul_node_47/adapter_weight_transpose output_name: model.layers.1.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 1 rank: 16 alpha: 32.0 - id: model.layers.10.mlp.down_proj component: model initializer: model.layers.10.mlp.down_proj.weight node_name: model/layers.10/mlp/down_proj/MatMul_node_297/adapter_weight_transpose output_name: model.layers.10.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.mlp.gate_proj component: model initializer: model.layers.10.mlp.gate_proj.weight node_name: model/layers.10/mlp/gate_proj/MatMul_node_291/adapter_weight_transpose output_name: model.layers.10.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.mlp.up_proj component: model initializer: model.layers.10.mlp.up_proj.weight node_name: model/layers.10/mlp/up_proj/MatMul_node_294/adapter_weight_transpose output_name: model.layers.10.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.self_attn.k_proj component: model initializer: model.layers.10.self_attn.k_proj.weight node_name: model/layers.10/self_attn/k_proj/MatMul_node_278/adapter_weight_transpose output_name: model.layers.10.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.self_attn.o_proj component: model initializer: model.layers.10.self_attn.o_proj.weight node_name: model/layers.10/self_attn/o_proj/MatMul_node_287/adapter_weight_transpose output_name: model.layers.10.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.self_attn.q_proj component: model initializer: model.layers.10.self_attn.q_proj.weight node_name: model/layers.10/self_attn/q_proj/MatMul_node_275/adapter_weight_transpose output_name: model.layers.10.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.10.self_attn.v_proj component: model initializer: model.layers.10.self_attn.v_proj.weight node_name: model/layers.10/self_attn/v_proj/MatMul_node_281/adapter_weight_transpose output_name: model.layers.10.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 10 rank: 16 alpha: 32.0 - id: model.layers.11.mlp.down_proj component: model initializer: model.layers.11.mlp.down_proj.weight node_name: model/layers.11/mlp/down_proj/MatMul_node_323/adapter_weight_transpose output_name: model.layers.11.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.mlp.gate_proj component: model initializer: model.layers.11.mlp.gate_proj.weight node_name: model/layers.11/mlp/gate_proj/MatMul_node_317/adapter_weight_transpose output_name: model.layers.11.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.mlp.up_proj component: model initializer: model.layers.11.mlp.up_proj.weight node_name: model/layers.11/mlp/up_proj/MatMul_node_320/adapter_weight_transpose output_name: model.layers.11.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.self_attn.k_proj component: model initializer: model.layers.11.self_attn.k_proj.weight node_name: model/layers.11/self_attn/k_proj/MatMul_node_304/adapter_weight_transpose output_name: model.layers.11.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.self_attn.o_proj component: model initializer: model.layers.11.self_attn.o_proj.weight node_name: model/layers.11/self_attn/o_proj/MatMul_node_313/adapter_weight_transpose output_name: model.layers.11.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.self_attn.q_proj component: model initializer: model.layers.11.self_attn.q_proj.weight node_name: model/layers.11/self_attn/q_proj/MatMul_node_301/adapter_weight_transpose output_name: model.layers.11.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.11.self_attn.v_proj component: model initializer: model.layers.11.self_attn.v_proj.weight node_name: model/layers.11/self_attn/v_proj/MatMul_node_307/adapter_weight_transpose output_name: model.layers.11.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 11 rank: 16 alpha: 32.0 - id: model.layers.12.mlp.down_proj component: model initializer: model.layers.12.mlp.down_proj.weight node_name: model/layers.12/mlp/down_proj/MatMul_node_349/adapter_weight_transpose output_name: model.layers.12.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.mlp.gate_proj component: model initializer: model.layers.12.mlp.gate_proj.weight node_name: model/layers.12/mlp/gate_proj/MatMul_node_343/adapter_weight_transpose output_name: model.layers.12.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.mlp.up_proj component: model initializer: model.layers.12.mlp.up_proj.weight node_name: model/layers.12/mlp/up_proj/MatMul_node_346/adapter_weight_transpose output_name: model.layers.12.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.self_attn.k_proj component: model initializer: model.layers.12.self_attn.k_proj.weight node_name: model/layers.12/self_attn/k_proj/MatMul_node_330/adapter_weight_transpose output_name: model.layers.12.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.self_attn.o_proj component: model initializer: model.layers.12.self_attn.o_proj.weight node_name: model/layers.12/self_attn/o_proj/MatMul_node_339/adapter_weight_transpose output_name: model.layers.12.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.self_attn.q_proj component: model initializer: model.layers.12.self_attn.q_proj.weight node_name: model/layers.12/self_attn/q_proj/MatMul_node_327/adapter_weight_transpose output_name: model.layers.12.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.12.self_attn.v_proj component: model initializer: model.layers.12.self_attn.v_proj.weight node_name: model/layers.12/self_attn/v_proj/MatMul_node_333/adapter_weight_transpose output_name: model.layers.12.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 12 rank: 16 alpha: 32.0 - id: model.layers.13.mlp.down_proj component: model initializer: model.layers.13.mlp.down_proj.weight node_name: model/layers.13/mlp/down_proj/MatMul_node_375/adapter_weight_transpose output_name: model.layers.13.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.mlp.gate_proj component: model initializer: model.layers.13.mlp.gate_proj.weight node_name: model/layers.13/mlp/gate_proj/MatMul_node_369/adapter_weight_transpose output_name: model.layers.13.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.mlp.up_proj component: model initializer: model.layers.13.mlp.up_proj.weight node_name: model/layers.13/mlp/up_proj/MatMul_node_372/adapter_weight_transpose output_name: model.layers.13.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.self_attn.k_proj component: model initializer: model.layers.13.self_attn.k_proj.weight node_name: model/layers.13/self_attn/k_proj/MatMul_node_356/adapter_weight_transpose output_name: model.layers.13.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.self_attn.o_proj component: model initializer: model.layers.13.self_attn.o_proj.weight node_name: model/layers.13/self_attn/o_proj/MatMul_node_365/adapter_weight_transpose output_name: model.layers.13.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.self_attn.q_proj component: model initializer: model.layers.13.self_attn.q_proj.weight node_name: model/layers.13/self_attn/q_proj/MatMul_node_353/adapter_weight_transpose output_name: model.layers.13.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.13.self_attn.v_proj component: model initializer: model.layers.13.self_attn.v_proj.weight node_name: model/layers.13/self_attn/v_proj/MatMul_node_359/adapter_weight_transpose output_name: model.layers.13.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 13 rank: 16 alpha: 32.0 - id: model.layers.14.mlp.down_proj component: model initializer: model.layers.14.mlp.down_proj.weight node_name: model/layers.14/mlp/down_proj/MatMul_node_401/adapter_weight_transpose output_name: model.layers.14.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.mlp.gate_proj component: model initializer: model.layers.14.mlp.gate_proj.weight node_name: model/layers.14/mlp/gate_proj/MatMul_node_395/adapter_weight_transpose output_name: model.layers.14.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.mlp.up_proj component: model initializer: model.layers.14.mlp.up_proj.weight node_name: model/layers.14/mlp/up_proj/MatMul_node_398/adapter_weight_transpose output_name: model.layers.14.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.self_attn.k_proj component: model initializer: model.layers.14.self_attn.k_proj.weight node_name: model/layers.14/self_attn/k_proj/MatMul_node_382/adapter_weight_transpose output_name: model.layers.14.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.self_attn.o_proj component: model initializer: model.layers.14.self_attn.o_proj.weight node_name: model/layers.14/self_attn/o_proj/MatMul_node_391/adapter_weight_transpose output_name: model.layers.14.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.self_attn.q_proj component: model initializer: model.layers.14.self_attn.q_proj.weight node_name: model/layers.14/self_attn/q_proj/MatMul_node_379/adapter_weight_transpose output_name: model.layers.14.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.14.self_attn.v_proj component: model initializer: model.layers.14.self_attn.v_proj.weight node_name: model/layers.14/self_attn/v_proj/MatMul_node_385/adapter_weight_transpose output_name: model.layers.14.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 14 rank: 16 alpha: 32.0 - id: model.layers.15.mlp.down_proj component: model initializer: model.layers.15.mlp.down_proj.weight node_name: model/layers.15/mlp/down_proj/MatMul_node_427/adapter_weight_transpose output_name: model.layers.15.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.mlp.gate_proj component: model initializer: model.layers.15.mlp.gate_proj.weight node_name: model/layers.15/mlp/gate_proj/MatMul_node_421/adapter_weight_transpose output_name: model.layers.15.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.mlp.up_proj component: model initializer: model.layers.15.mlp.up_proj.weight node_name: model/layers.15/mlp/up_proj/MatMul_node_424/adapter_weight_transpose output_name: model.layers.15.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.self_attn.k_proj component: model initializer: model.layers.15.self_attn.k_proj.weight node_name: model/layers.15/self_attn/k_proj/MatMul_node_408/adapter_weight_transpose output_name: model.layers.15.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.self_attn.o_proj component: model initializer: model.layers.15.self_attn.o_proj.weight node_name: model/layers.15/self_attn/o_proj/MatMul_node_417/adapter_weight_transpose output_name: model.layers.15.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.self_attn.q_proj component: model initializer: model.layers.15.self_attn.q_proj.weight node_name: model/layers.15/self_attn/q_proj/MatMul_node_405/adapter_weight_transpose output_name: model.layers.15.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.15.self_attn.v_proj component: model initializer: model.layers.15.self_attn.v_proj.weight node_name: model/layers.15/self_attn/v_proj/MatMul_node_411/adapter_weight_transpose output_name: model.layers.15.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 15 rank: 16 alpha: 32.0 - id: model.layers.16.mlp.down_proj component: model initializer: model.layers.16.mlp.down_proj.weight node_name: model/layers.16/mlp/down_proj/MatMul_node_453/adapter_weight_transpose output_name: model.layers.16.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.mlp.gate_proj component: model initializer: model.layers.16.mlp.gate_proj.weight node_name: model/layers.16/mlp/gate_proj/MatMul_node_447/adapter_weight_transpose output_name: model.layers.16.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.mlp.up_proj component: model initializer: model.layers.16.mlp.up_proj.weight node_name: model/layers.16/mlp/up_proj/MatMul_node_450/adapter_weight_transpose output_name: model.layers.16.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.self_attn.k_proj component: model initializer: model.layers.16.self_attn.k_proj.weight node_name: model/layers.16/self_attn/k_proj/MatMul_node_434/adapter_weight_transpose output_name: model.layers.16.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.self_attn.o_proj component: model initializer: model.layers.16.self_attn.o_proj.weight node_name: model/layers.16/self_attn/o_proj/MatMul_node_443/adapter_weight_transpose output_name: model.layers.16.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.self_attn.q_proj component: model initializer: model.layers.16.self_attn.q_proj.weight node_name: model/layers.16/self_attn/q_proj/MatMul_node_431/adapter_weight_transpose output_name: model.layers.16.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.16.self_attn.v_proj component: model initializer: model.layers.16.self_attn.v_proj.weight node_name: model/layers.16/self_attn/v_proj/MatMul_node_437/adapter_weight_transpose output_name: model.layers.16.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 16 rank: 16 alpha: 32.0 - id: model.layers.17.mlp.down_proj component: model initializer: model.layers.17.mlp.down_proj.weight node_name: model/layers.17/mlp/down_proj/MatMul_node_479/adapter_weight_transpose output_name: model.layers.17.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.mlp.gate_proj component: model initializer: model.layers.17.mlp.gate_proj.weight node_name: model/layers.17/mlp/gate_proj/MatMul_node_473/adapter_weight_transpose output_name: model.layers.17.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.mlp.up_proj component: model initializer: model.layers.17.mlp.up_proj.weight node_name: model/layers.17/mlp/up_proj/MatMul_node_476/adapter_weight_transpose output_name: model.layers.17.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.self_attn.k_proj component: model initializer: model.layers.17.self_attn.k_proj.weight node_name: model/layers.17/self_attn/k_proj/MatMul_node_460/adapter_weight_transpose output_name: model.layers.17.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.self_attn.o_proj component: model initializer: model.layers.17.self_attn.o_proj.weight node_name: model/layers.17/self_attn/o_proj/MatMul_node_469/adapter_weight_transpose output_name: model.layers.17.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.self_attn.q_proj component: model initializer: model.layers.17.self_attn.q_proj.weight node_name: model/layers.17/self_attn/q_proj/MatMul_node_457/adapter_weight_transpose output_name: model.layers.17.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.17.self_attn.v_proj component: model initializer: model.layers.17.self_attn.v_proj.weight node_name: model/layers.17/self_attn/v_proj/MatMul_node_463/adapter_weight_transpose output_name: model.layers.17.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 17 rank: 16 alpha: 32.0 - id: model.layers.18.mlp.down_proj component: model initializer: model.layers.18.mlp.down_proj.weight node_name: model/layers.18/mlp/down_proj/MatMul_node_505/adapter_weight_transpose output_name: model.layers.18.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.mlp.gate_proj component: model initializer: model.layers.18.mlp.gate_proj.weight node_name: model/layers.18/mlp/gate_proj/MatMul_node_499/adapter_weight_transpose output_name: model.layers.18.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.mlp.up_proj component: model initializer: model.layers.18.mlp.up_proj.weight node_name: model/layers.18/mlp/up_proj/MatMul_node_502/adapter_weight_transpose output_name: model.layers.18.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.self_attn.k_proj component: model initializer: model.layers.18.self_attn.k_proj.weight node_name: model/layers.18/self_attn/k_proj/MatMul_node_486/adapter_weight_transpose output_name: model.layers.18.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.self_attn.o_proj component: model initializer: model.layers.18.self_attn.o_proj.weight node_name: model/layers.18/self_attn/o_proj/MatMul_node_495/adapter_weight_transpose output_name: model.layers.18.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.self_attn.q_proj component: model initializer: model.layers.18.self_attn.q_proj.weight node_name: model/layers.18/self_attn/q_proj/MatMul_node_483/adapter_weight_transpose output_name: model.layers.18.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.18.self_attn.v_proj component: model initializer: model.layers.18.self_attn.v_proj.weight node_name: model/layers.18/self_attn/v_proj/MatMul_node_489/adapter_weight_transpose output_name: model.layers.18.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 18 rank: 16 alpha: 32.0 - id: model.layers.19.mlp.down_proj component: model initializer: model.layers.19.mlp.down_proj.weight node_name: model/layers.19/mlp/down_proj/MatMul_node_531/adapter_weight_transpose output_name: model.layers.19.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.mlp.gate_proj component: model initializer: model.layers.19.mlp.gate_proj.weight node_name: model/layers.19/mlp/gate_proj/MatMul_node_525/adapter_weight_transpose output_name: model.layers.19.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.mlp.up_proj component: model initializer: model.layers.19.mlp.up_proj.weight node_name: model/layers.19/mlp/up_proj/MatMul_node_528/adapter_weight_transpose output_name: model.layers.19.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.self_attn.k_proj component: model initializer: model.layers.19.self_attn.k_proj.weight node_name: model/layers.19/self_attn/k_proj/MatMul_node_512/adapter_weight_transpose output_name: model.layers.19.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.self_attn.o_proj component: model initializer: model.layers.19.self_attn.o_proj.weight node_name: model/layers.19/self_attn/o_proj/MatMul_node_521/adapter_weight_transpose output_name: model.layers.19.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.self_attn.q_proj component: model initializer: model.layers.19.self_attn.q_proj.weight node_name: model/layers.19/self_attn/q_proj/MatMul_node_509/adapter_weight_transpose output_name: model.layers.19.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.19.self_attn.v_proj component: model initializer: model.layers.19.self_attn.v_proj.weight node_name: model/layers.19/self_attn/v_proj/MatMul_node_515/adapter_weight_transpose output_name: model.layers.19.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 19 rank: 16 alpha: 32.0 - id: model.layers.2.mlp.down_proj component: model initializer: model.layers.2.mlp.down_proj.weight node_name: model/layers.2/mlp/down_proj/MatMul_node_89/adapter_weight_transpose output_name: model.layers.2.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.mlp.gate_proj component: model initializer: model.layers.2.mlp.gate_proj.weight node_name: model/layers.2/mlp/gate_proj/MatMul_node_83/adapter_weight_transpose output_name: model.layers.2.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.mlp.up_proj component: model initializer: model.layers.2.mlp.up_proj.weight node_name: model/layers.2/mlp/up_proj/MatMul_node_86/adapter_weight_transpose output_name: model.layers.2.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.self_attn.k_proj component: model initializer: model.layers.2.self_attn.k_proj.weight node_name: model/layers.2/self_attn/k_proj/MatMul_node_70/adapter_weight_transpose output_name: model.layers.2.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.self_attn.o_proj component: model initializer: model.layers.2.self_attn.o_proj.weight node_name: model/layers.2/self_attn/o_proj/MatMul_node_79/adapter_weight_transpose output_name: model.layers.2.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.self_attn.q_proj component: model initializer: model.layers.2.self_attn.q_proj.weight node_name: model/layers.2/self_attn/q_proj/MatMul_node_67/adapter_weight_transpose output_name: model.layers.2.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.2.self_attn.v_proj component: model initializer: model.layers.2.self_attn.v_proj.weight node_name: model/layers.2/self_attn/v_proj/MatMul_node_73/adapter_weight_transpose output_name: model.layers.2.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 2 rank: 16 alpha: 32.0 - id: model.layers.20.mlp.down_proj component: model initializer: model.layers.20.mlp.down_proj.weight node_name: model/layers.20/mlp/down_proj/MatMul_node_557/adapter_weight_transpose output_name: model.layers.20.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.mlp.gate_proj component: model initializer: model.layers.20.mlp.gate_proj.weight node_name: model/layers.20/mlp/gate_proj/MatMul_node_551/adapter_weight_transpose output_name: model.layers.20.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.mlp.up_proj component: model initializer: model.layers.20.mlp.up_proj.weight node_name: model/layers.20/mlp/up_proj/MatMul_node_554/adapter_weight_transpose output_name: model.layers.20.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.self_attn.k_proj component: model initializer: model.layers.20.self_attn.k_proj.weight node_name: model/layers.20/self_attn/k_proj/MatMul_node_538/adapter_weight_transpose output_name: model.layers.20.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.self_attn.o_proj component: model initializer: model.layers.20.self_attn.o_proj.weight node_name: model/layers.20/self_attn/o_proj/MatMul_node_547/adapter_weight_transpose output_name: model.layers.20.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.self_attn.q_proj component: model initializer: model.layers.20.self_attn.q_proj.weight node_name: model/layers.20/self_attn/q_proj/MatMul_node_535/adapter_weight_transpose output_name: model.layers.20.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.20.self_attn.v_proj component: model initializer: model.layers.20.self_attn.v_proj.weight node_name: model/layers.20/self_attn/v_proj/MatMul_node_541/adapter_weight_transpose output_name: model.layers.20.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 20 rank: 16 alpha: 32.0 - id: model.layers.21.mlp.down_proj component: model initializer: model.layers.21.mlp.down_proj.weight node_name: model/layers.21/mlp/down_proj/MatMul_node_583/adapter_weight_transpose output_name: model.layers.21.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.mlp.gate_proj component: model initializer: model.layers.21.mlp.gate_proj.weight node_name: model/layers.21/mlp/gate_proj/MatMul_node_577/adapter_weight_transpose output_name: model.layers.21.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.mlp.up_proj component: model initializer: model.layers.21.mlp.up_proj.weight node_name: model/layers.21/mlp/up_proj/MatMul_node_580/adapter_weight_transpose output_name: model.layers.21.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.self_attn.k_proj component: model initializer: model.layers.21.self_attn.k_proj.weight node_name: model/layers.21/self_attn/k_proj/MatMul_node_564/adapter_weight_transpose output_name: model.layers.21.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.self_attn.o_proj component: model initializer: model.layers.21.self_attn.o_proj.weight node_name: model/layers.21/self_attn/o_proj/MatMul_node_573/adapter_weight_transpose output_name: model.layers.21.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.self_attn.q_proj component: model initializer: model.layers.21.self_attn.q_proj.weight node_name: model/layers.21/self_attn/q_proj/MatMul_node_561/adapter_weight_transpose output_name: model.layers.21.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.21.self_attn.v_proj component: model initializer: model.layers.21.self_attn.v_proj.weight node_name: model/layers.21/self_attn/v_proj/MatMul_node_567/adapter_weight_transpose output_name: model.layers.21.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 21 rank: 16 alpha: 32.0 - id: model.layers.22.mlp.down_proj component: model initializer: model.layers.22.mlp.down_proj.weight node_name: model/layers.22/mlp/down_proj/MatMul_node_609/adapter_weight_transpose output_name: model.layers.22.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.mlp.gate_proj component: model initializer: model.layers.22.mlp.gate_proj.weight node_name: model/layers.22/mlp/gate_proj/MatMul_node_603/adapter_weight_transpose output_name: model.layers.22.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.mlp.up_proj component: model initializer: model.layers.22.mlp.up_proj.weight node_name: model/layers.22/mlp/up_proj/MatMul_node_606/adapter_weight_transpose output_name: model.layers.22.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.self_attn.k_proj component: model initializer: model.layers.22.self_attn.k_proj.weight node_name: model/layers.22/self_attn/k_proj/MatMul_node_590/adapter_weight_transpose output_name: model.layers.22.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.self_attn.o_proj component: model initializer: model.layers.22.self_attn.o_proj.weight node_name: model/layers.22/self_attn/o_proj/MatMul_node_599/adapter_weight_transpose output_name: model.layers.22.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.self_attn.q_proj component: model initializer: model.layers.22.self_attn.q_proj.weight node_name: model/layers.22/self_attn/q_proj/MatMul_node_587/adapter_weight_transpose output_name: model.layers.22.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.22.self_attn.v_proj component: model initializer: model.layers.22.self_attn.v_proj.weight node_name: model/layers.22/self_attn/v_proj/MatMul_node_593/adapter_weight_transpose output_name: model.layers.22.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 22 rank: 16 alpha: 32.0 - id: model.layers.23.mlp.down_proj component: model initializer: model.layers.23.mlp.down_proj.weight node_name: model/layers.23/mlp/down_proj/MatMul_node_635/adapter_weight_transpose output_name: model.layers.23.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.mlp.gate_proj component: model initializer: model.layers.23.mlp.gate_proj.weight node_name: model/layers.23/mlp/gate_proj/MatMul_node_629/adapter_weight_transpose output_name: model.layers.23.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.mlp.up_proj component: model initializer: model.layers.23.mlp.up_proj.weight node_name: model/layers.23/mlp/up_proj/MatMul_node_632/adapter_weight_transpose output_name: model.layers.23.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.self_attn.k_proj component: model initializer: model.layers.23.self_attn.k_proj.weight node_name: model/layers.23/self_attn/k_proj/MatMul_node_616/adapter_weight_transpose output_name: model.layers.23.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.self_attn.o_proj component: model initializer: model.layers.23.self_attn.o_proj.weight node_name: model/layers.23/self_attn/o_proj/MatMul_node_625/adapter_weight_transpose output_name: model.layers.23.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.self_attn.q_proj component: model initializer: model.layers.23.self_attn.q_proj.weight node_name: model/layers.23/self_attn/q_proj/MatMul_node_613/adapter_weight_transpose output_name: model.layers.23.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.23.self_attn.v_proj component: model initializer: model.layers.23.self_attn.v_proj.weight node_name: model/layers.23/self_attn/v_proj/MatMul_node_619/adapter_weight_transpose output_name: model.layers.23.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 23 rank: 16 alpha: 32.0 - id: model.layers.24.mlp.down_proj component: model initializer: model.layers.24.mlp.down_proj.weight node_name: model/layers.24/mlp/down_proj/MatMul_node_661/adapter_weight_transpose output_name: model.layers.24.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.mlp.gate_proj component: model initializer: model.layers.24.mlp.gate_proj.weight node_name: model/layers.24/mlp/gate_proj/MatMul_node_655/adapter_weight_transpose output_name: model.layers.24.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.mlp.up_proj component: model initializer: model.layers.24.mlp.up_proj.weight node_name: model/layers.24/mlp/up_proj/MatMul_node_658/adapter_weight_transpose output_name: model.layers.24.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.self_attn.k_proj component: model initializer: model.layers.24.self_attn.k_proj.weight node_name: model/layers.24/self_attn/k_proj/MatMul_node_642/adapter_weight_transpose output_name: model.layers.24.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.self_attn.o_proj component: model initializer: model.layers.24.self_attn.o_proj.weight node_name: model/layers.24/self_attn/o_proj/MatMul_node_651/adapter_weight_transpose output_name: model.layers.24.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.self_attn.q_proj component: model initializer: model.layers.24.self_attn.q_proj.weight node_name: model/layers.24/self_attn/q_proj/MatMul_node_639/adapter_weight_transpose output_name: model.layers.24.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.24.self_attn.v_proj component: model initializer: model.layers.24.self_attn.v_proj.weight node_name: model/layers.24/self_attn/v_proj/MatMul_node_645/adapter_weight_transpose output_name: model.layers.24.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 24 rank: 16 alpha: 32.0 - id: model.layers.25.mlp.down_proj component: model initializer: model.layers.25.mlp.down_proj.weight node_name: model/layers.25/mlp/down_proj/MatMul_node_687/adapter_weight_transpose output_name: model.layers.25.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.mlp.gate_proj component: model initializer: model.layers.25.mlp.gate_proj.weight node_name: model/layers.25/mlp/gate_proj/MatMul_node_681/adapter_weight_transpose output_name: model.layers.25.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.mlp.up_proj component: model initializer: model.layers.25.mlp.up_proj.weight node_name: model/layers.25/mlp/up_proj/MatMul_node_684/adapter_weight_transpose output_name: model.layers.25.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.self_attn.k_proj component: model initializer: model.layers.25.self_attn.k_proj.weight node_name: model/layers.25/self_attn/k_proj/MatMul_node_668/adapter_weight_transpose output_name: model.layers.25.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.self_attn.o_proj component: model initializer: model.layers.25.self_attn.o_proj.weight node_name: model/layers.25/self_attn/o_proj/MatMul_node_677/adapter_weight_transpose output_name: model.layers.25.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.self_attn.q_proj component: model initializer: model.layers.25.self_attn.q_proj.weight node_name: model/layers.25/self_attn/q_proj/MatMul_node_665/adapter_weight_transpose output_name: model.layers.25.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.25.self_attn.v_proj component: model initializer: model.layers.25.self_attn.v_proj.weight node_name: model/layers.25/self_attn/v_proj/MatMul_node_671/adapter_weight_transpose output_name: model.layers.25.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 25 rank: 16 alpha: 32.0 - id: model.layers.26.mlp.down_proj component: model initializer: model.layers.26.mlp.down_proj.weight node_name: model/layers.26/mlp/down_proj/MatMul_node_713/adapter_weight_transpose output_name: model.layers.26.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.mlp.gate_proj component: model initializer: model.layers.26.mlp.gate_proj.weight node_name: model/layers.26/mlp/gate_proj/MatMul_node_707/adapter_weight_transpose output_name: model.layers.26.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.mlp.up_proj component: model initializer: model.layers.26.mlp.up_proj.weight node_name: model/layers.26/mlp/up_proj/MatMul_node_710/adapter_weight_transpose output_name: model.layers.26.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.self_attn.k_proj component: model initializer: model.layers.26.self_attn.k_proj.weight node_name: model/layers.26/self_attn/k_proj/MatMul_node_694/adapter_weight_transpose output_name: model.layers.26.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.self_attn.o_proj component: model initializer: model.layers.26.self_attn.o_proj.weight node_name: model/layers.26/self_attn/o_proj/MatMul_node_703/adapter_weight_transpose output_name: model.layers.26.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.self_attn.q_proj component: model initializer: model.layers.26.self_attn.q_proj.weight node_name: model/layers.26/self_attn/q_proj/MatMul_node_691/adapter_weight_transpose output_name: model.layers.26.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.26.self_attn.v_proj component: model initializer: model.layers.26.self_attn.v_proj.weight node_name: model/layers.26/self_attn/v_proj/MatMul_node_697/adapter_weight_transpose output_name: model.layers.26.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 26 rank: 16 alpha: 32.0 - id: model.layers.27.mlp.down_proj component: model initializer: model.layers.27.mlp.down_proj.weight node_name: model/layers.27/mlp/down_proj/MatMul_node_739/adapter_weight_transpose output_name: model.layers.27.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.mlp.gate_proj component: model initializer: model.layers.27.mlp.gate_proj.weight node_name: model/layers.27/mlp/gate_proj/MatMul_node_733/adapter_weight_transpose output_name: model.layers.27.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.mlp.up_proj component: model initializer: model.layers.27.mlp.up_proj.weight node_name: model/layers.27/mlp/up_proj/MatMul_node_736/adapter_weight_transpose output_name: model.layers.27.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.self_attn.k_proj component: model initializer: model.layers.27.self_attn.k_proj.weight node_name: model/layers.27/self_attn/k_proj/MatMul_node_720/adapter_weight_transpose output_name: model.layers.27.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.self_attn.o_proj component: model initializer: model.layers.27.self_attn.o_proj.weight node_name: model/layers.27/self_attn/o_proj/MatMul_node_729/adapter_weight_transpose output_name: model.layers.27.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.self_attn.q_proj component: model initializer: model.layers.27.self_attn.q_proj.weight node_name: model/layers.27/self_attn/q_proj/MatMul_node_717/adapter_weight_transpose output_name: model.layers.27.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.27.self_attn.v_proj component: model initializer: model.layers.27.self_attn.v_proj.weight node_name: model/layers.27/self_attn/v_proj/MatMul_node_723/adapter_weight_transpose output_name: model.layers.27.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 27 rank: 16 alpha: 32.0 - id: model.layers.3.mlp.down_proj component: model initializer: model.layers.3.mlp.down_proj.weight node_name: model/layers.3/mlp/down_proj/MatMul_node_115/adapter_weight_transpose output_name: model.layers.3.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.mlp.gate_proj component: model initializer: model.layers.3.mlp.gate_proj.weight node_name: model/layers.3/mlp/gate_proj/MatMul_node_109/adapter_weight_transpose output_name: model.layers.3.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.mlp.up_proj component: model initializer: model.layers.3.mlp.up_proj.weight node_name: model/layers.3/mlp/up_proj/MatMul_node_112/adapter_weight_transpose output_name: model.layers.3.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.self_attn.k_proj component: model initializer: model.layers.3.self_attn.k_proj.weight node_name: model/layers.3/self_attn/k_proj/MatMul_node_96/adapter_weight_transpose output_name: model.layers.3.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.self_attn.o_proj component: model initializer: model.layers.3.self_attn.o_proj.weight node_name: model/layers.3/self_attn/o_proj/MatMul_node_105/adapter_weight_transpose output_name: model.layers.3.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.self_attn.q_proj component: model initializer: model.layers.3.self_attn.q_proj.weight node_name: model/layers.3/self_attn/q_proj/MatMul_node_93/adapter_weight_transpose output_name: model.layers.3.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.3.self_attn.v_proj component: model initializer: model.layers.3.self_attn.v_proj.weight node_name: model/layers.3/self_attn/v_proj/MatMul_node_99/adapter_weight_transpose output_name: model.layers.3.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 3 rank: 16 alpha: 32.0 - id: model.layers.4.mlp.down_proj component: model initializer: model.layers.4.mlp.down_proj.weight node_name: model/layers.4/mlp/down_proj/MatMul_node_141/adapter_weight_transpose output_name: model.layers.4.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.mlp.gate_proj component: model initializer: model.layers.4.mlp.gate_proj.weight node_name: model/layers.4/mlp/gate_proj/MatMul_node_135/adapter_weight_transpose output_name: model.layers.4.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.mlp.up_proj component: model initializer: model.layers.4.mlp.up_proj.weight node_name: model/layers.4/mlp/up_proj/MatMul_node_138/adapter_weight_transpose output_name: model.layers.4.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.self_attn.k_proj component: model initializer: model.layers.4.self_attn.k_proj.weight node_name: model/layers.4/self_attn/k_proj/MatMul_node_122/adapter_weight_transpose output_name: model.layers.4.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.self_attn.o_proj component: model initializer: model.layers.4.self_attn.o_proj.weight node_name: model/layers.4/self_attn/o_proj/MatMul_node_131/adapter_weight_transpose output_name: model.layers.4.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.self_attn.q_proj component: model initializer: model.layers.4.self_attn.q_proj.weight node_name: model/layers.4/self_attn/q_proj/MatMul_node_119/adapter_weight_transpose output_name: model.layers.4.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.4.self_attn.v_proj component: model initializer: model.layers.4.self_attn.v_proj.weight node_name: model/layers.4/self_attn/v_proj/MatMul_node_125/adapter_weight_transpose output_name: model.layers.4.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 4 rank: 16 alpha: 32.0 - id: model.layers.5.mlp.down_proj component: model initializer: model.layers.5.mlp.down_proj.weight node_name: model/layers.5/mlp/down_proj/MatMul_node_167/adapter_weight_transpose output_name: model.layers.5.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.mlp.gate_proj component: model initializer: model.layers.5.mlp.gate_proj.weight node_name: model/layers.5/mlp/gate_proj/MatMul_node_161/adapter_weight_transpose output_name: model.layers.5.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.mlp.up_proj component: model initializer: model.layers.5.mlp.up_proj.weight node_name: model/layers.5/mlp/up_proj/MatMul_node_164/adapter_weight_transpose output_name: model.layers.5.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.self_attn.k_proj component: model initializer: model.layers.5.self_attn.k_proj.weight node_name: model/layers.5/self_attn/k_proj/MatMul_node_148/adapter_weight_transpose output_name: model.layers.5.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.self_attn.o_proj component: model initializer: model.layers.5.self_attn.o_proj.weight node_name: model/layers.5/self_attn/o_proj/MatMul_node_157/adapter_weight_transpose output_name: model.layers.5.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.self_attn.q_proj component: model initializer: model.layers.5.self_attn.q_proj.weight node_name: model/layers.5/self_attn/q_proj/MatMul_node_145/adapter_weight_transpose output_name: model.layers.5.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.5.self_attn.v_proj component: model initializer: model.layers.5.self_attn.v_proj.weight node_name: model/layers.5/self_attn/v_proj/MatMul_node_151/adapter_weight_transpose output_name: model.layers.5.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 5 rank: 16 alpha: 32.0 - id: model.layers.6.mlp.down_proj component: model initializer: model.layers.6.mlp.down_proj.weight node_name: model/layers.6/mlp/down_proj/MatMul_node_193/adapter_weight_transpose output_name: model.layers.6.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.mlp.gate_proj component: model initializer: model.layers.6.mlp.gate_proj.weight node_name: model/layers.6/mlp/gate_proj/MatMul_node_187/adapter_weight_transpose output_name: model.layers.6.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.mlp.up_proj component: model initializer: model.layers.6.mlp.up_proj.weight node_name: model/layers.6/mlp/up_proj/MatMul_node_190/adapter_weight_transpose output_name: model.layers.6.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.self_attn.k_proj component: model initializer: model.layers.6.self_attn.k_proj.weight node_name: model/layers.6/self_attn/k_proj/MatMul_node_174/adapter_weight_transpose output_name: model.layers.6.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.self_attn.o_proj component: model initializer: model.layers.6.self_attn.o_proj.weight node_name: model/layers.6/self_attn/o_proj/MatMul_node_183/adapter_weight_transpose output_name: model.layers.6.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.self_attn.q_proj component: model initializer: model.layers.6.self_attn.q_proj.weight node_name: model/layers.6/self_attn/q_proj/MatMul_node_171/adapter_weight_transpose output_name: model.layers.6.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.6.self_attn.v_proj component: model initializer: model.layers.6.self_attn.v_proj.weight node_name: model/layers.6/self_attn/v_proj/MatMul_node_177/adapter_weight_transpose output_name: model.layers.6.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 6 rank: 16 alpha: 32.0 - id: model.layers.7.mlp.down_proj component: model initializer: model.layers.7.mlp.down_proj.weight node_name: model/layers.7/mlp/down_proj/MatMul_node_219/adapter_weight_transpose output_name: model.layers.7.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.mlp.gate_proj component: model initializer: model.layers.7.mlp.gate_proj.weight node_name: model/layers.7/mlp/gate_proj/MatMul_node_213/adapter_weight_transpose output_name: model.layers.7.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.mlp.up_proj component: model initializer: model.layers.7.mlp.up_proj.weight node_name: model/layers.7/mlp/up_proj/MatMul_node_216/adapter_weight_transpose output_name: model.layers.7.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.self_attn.k_proj component: model initializer: model.layers.7.self_attn.k_proj.weight node_name: model/layers.7/self_attn/k_proj/MatMul_node_200/adapter_weight_transpose output_name: model.layers.7.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.self_attn.o_proj component: model initializer: model.layers.7.self_attn.o_proj.weight node_name: model/layers.7/self_attn/o_proj/MatMul_node_209/adapter_weight_transpose output_name: model.layers.7.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.self_attn.q_proj component: model initializer: model.layers.7.self_attn.q_proj.weight node_name: model/layers.7/self_attn/q_proj/MatMul_node_197/adapter_weight_transpose output_name: model.layers.7.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.7.self_attn.v_proj component: model initializer: model.layers.7.self_attn.v_proj.weight node_name: model/layers.7/self_attn/v_proj/MatMul_node_203/adapter_weight_transpose output_name: model.layers.7.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 7 rank: 16 alpha: 32.0 - id: model.layers.8.mlp.down_proj component: model initializer: model.layers.8.mlp.down_proj.weight node_name: model/layers.8/mlp/down_proj/MatMul_node_245/adapter_weight_transpose output_name: model.layers.8.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.mlp.gate_proj component: model initializer: model.layers.8.mlp.gate_proj.weight node_name: model/layers.8/mlp/gate_proj/MatMul_node_239/adapter_weight_transpose output_name: model.layers.8.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.mlp.up_proj component: model initializer: model.layers.8.mlp.up_proj.weight node_name: model/layers.8/mlp/up_proj/MatMul_node_242/adapter_weight_transpose output_name: model.layers.8.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.self_attn.k_proj component: model initializer: model.layers.8.self_attn.k_proj.weight node_name: model/layers.8/self_attn/k_proj/MatMul_node_226/adapter_weight_transpose output_name: model.layers.8.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.self_attn.o_proj component: model initializer: model.layers.8.self_attn.o_proj.weight node_name: model/layers.8/self_attn/o_proj/MatMul_node_235/adapter_weight_transpose output_name: model.layers.8.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.self_attn.q_proj component: model initializer: model.layers.8.self_attn.q_proj.weight node_name: model/layers.8/self_attn/q_proj/MatMul_node_223/adapter_weight_transpose output_name: model.layers.8.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.8.self_attn.v_proj component: model initializer: model.layers.8.self_attn.v_proj.weight node_name: model/layers.8/self_attn/v_proj/MatMul_node_229/adapter_weight_transpose output_name: model.layers.8.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 8 rank: 16 alpha: 32.0 - id: model.layers.9.mlp.down_proj component: model initializer: model.layers.9.mlp.down_proj.weight node_name: model/layers.9/mlp/down_proj/MatMul_node_271/adapter_weight_transpose output_name: model.layers.9.mlp.down_proj.weight.transposed activation_dtype: float16 input_features: 8960 output_features: 1536 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.mlp.gate_proj component: model initializer: model.layers.9.mlp.gate_proj.weight node_name: model/layers.9/mlp/gate_proj/MatMul_node_265/adapter_weight_transpose output_name: model.layers.9.mlp.gate_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.mlp.up_proj component: model initializer: model.layers.9.mlp.up_proj.weight node_name: model/layers.9/mlp/up_proj/MatMul_node_268/adapter_weight_transpose output_name: model.layers.9.mlp.up_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 8960 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.self_attn.k_proj component: model initializer: model.layers.9.self_attn.k_proj.weight node_name: model/layers.9/self_attn/k_proj/MatMul_node_252/adapter_weight_transpose output_name: model.layers.9.self_attn.k_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.self_attn.o_proj component: model initializer: model.layers.9.self_attn.o_proj.weight node_name: model/layers.9/self_attn/o_proj/MatMul_node_261/adapter_weight_transpose output_name: model.layers.9.self_attn.o_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.self_attn.q_proj component: model initializer: model.layers.9.self_attn.q_proj.weight node_name: model/layers.9/self_attn/q_proj/MatMul_node_249/adapter_weight_transpose output_name: model.layers.9.self_attn.q_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 1536 layer_index: 9 rank: 16 alpha: 32.0 - id: model.layers.9.self_attn.v_proj component: model initializer: model.layers.9.self_attn.v_proj.weight node_name: model/layers.9/self_attn/v_proj/MatMul_node_255/adapter_weight_transpose output_name: model.layers.9.self_attn.v_proj.weight.transposed activation_dtype: float16 input_features: 1536 output_features: 256 layer_index: 9 rank: 16 alpha: 32.0 discovery_fallback: disabled selection: segments: request.adapter_segments adapter_counts: request.adapter_counts scales: request.adapter_scales max_adapters: 2 application_capability: onnx-genai.adapters@1 portable_fallback: true cache: max_entries: 2 eviction: lru planning: bucket_by_adapter_set: true stable_buffers: true invalidate_capture_on_eviction: true artifacts: code-lora-r16: index: 0 identity: code-lora-r16 version: '1' rank: 16 alpha: 32.0 dtype: float16 provenance: producer: mobius source: Qwen/Qwen2.5-1.5B-Instruct weights: - location: adapters/code-lora-r16/adapter.json loader_capability: onnx-genai.adapters.json@1 scale_encoding: alpha_over_rank format: json - location: adapters/code-lora-r16/adapter_model.safetensors loader_capability: onnx-genai.adapters.hf-peft@1 config_location: adapters/code-lora-r16/adapter_config.json scale_encoding: alpha_over_rank format: hf_peft bindings: - target: model.layers.0.mlp.down_proj weight_key: base_model.model.model.layers.0.mlp.down_proj - target: model.layers.0.mlp.gate_proj weight_key: base_model.model.model.layers.0.mlp.gate_proj - target: model.layers.0.mlp.up_proj weight_key: base_model.model.model.layers.0.mlp.up_proj - target: model.layers.0.self_attn.k_proj weight_key: base_model.model.model.layers.0.self_attn.k_proj - target: model.layers.0.self_attn.o_proj weight_key: base_model.model.model.layers.0.self_attn.o_proj - target: model.layers.0.self_attn.q_proj weight_key: base_model.model.model.layers.0.self_attn.q_proj - target: model.layers.0.self_attn.v_proj weight_key: base_model.model.model.layers.0.self_attn.v_proj - target: model.layers.1.mlp.down_proj weight_key: base_model.model.model.layers.1.mlp.down_proj - target: model.layers.1.mlp.gate_proj weight_key: base_model.model.model.layers.1.mlp.gate_proj - target: model.layers.1.mlp.up_proj weight_key: base_model.model.model.layers.1.mlp.up_proj - target: model.layers.1.self_attn.k_proj weight_key: base_model.model.model.layers.1.self_attn.k_proj - target: model.layers.1.self_attn.o_proj weight_key: base_model.model.model.layers.1.self_attn.o_proj - target: model.layers.1.self_attn.q_proj weight_key: base_model.model.model.layers.1.self_attn.q_proj - target: model.layers.1.self_attn.v_proj weight_key: base_model.model.model.layers.1.self_attn.v_proj - target: model.layers.10.mlp.down_proj weight_key: base_model.model.model.layers.10.mlp.down_proj - target: model.layers.10.mlp.gate_proj weight_key: base_model.model.model.layers.10.mlp.gate_proj - target: model.layers.10.mlp.up_proj weight_key: base_model.model.model.layers.10.mlp.up_proj - target: model.layers.10.self_attn.k_proj weight_key: base_model.model.model.layers.10.self_attn.k_proj - target: model.layers.10.self_attn.o_proj weight_key: base_model.model.model.layers.10.self_attn.o_proj - target: model.layers.10.self_attn.q_proj weight_key: base_model.model.model.layers.10.self_attn.q_proj - target: model.layers.10.self_attn.v_proj weight_key: base_model.model.model.layers.10.self_attn.v_proj - target: model.layers.11.mlp.down_proj weight_key: base_model.model.model.layers.11.mlp.down_proj - target: model.layers.11.mlp.gate_proj weight_key: base_model.model.model.layers.11.mlp.gate_proj - target: model.layers.11.mlp.up_proj weight_key: base_model.model.model.layers.11.mlp.up_proj - target: model.layers.11.self_attn.k_proj weight_key: base_model.model.model.layers.11.self_attn.k_proj - target: model.layers.11.self_attn.o_proj weight_key: base_model.model.model.layers.11.self_attn.o_proj - target: model.layers.11.self_attn.q_proj weight_key: base_model.model.model.layers.11.self_attn.q_proj - target: model.layers.11.self_attn.v_proj weight_key: base_model.model.model.layers.11.self_attn.v_proj - target: model.layers.12.mlp.down_proj weight_key: base_model.model.model.layers.12.mlp.down_proj - target: model.layers.12.mlp.gate_proj weight_key: base_model.model.model.layers.12.mlp.gate_proj - target: model.layers.12.mlp.up_proj weight_key: base_model.model.model.layers.12.mlp.up_proj - target: model.layers.12.self_attn.k_proj weight_key: base_model.model.model.layers.12.self_attn.k_proj - target: model.layers.12.self_attn.o_proj weight_key: base_model.model.model.layers.12.self_attn.o_proj - target: model.layers.12.self_attn.q_proj weight_key: base_model.model.model.layers.12.self_attn.q_proj - target: model.layers.12.self_attn.v_proj weight_key: base_model.model.model.layers.12.self_attn.v_proj - target: model.layers.13.mlp.down_proj weight_key: base_model.model.model.layers.13.mlp.down_proj - target: model.layers.13.mlp.gate_proj weight_key: base_model.model.model.layers.13.mlp.gate_proj - target: model.layers.13.mlp.up_proj weight_key: base_model.model.model.layers.13.mlp.up_proj - target: model.layers.13.self_attn.k_proj weight_key: base_model.model.model.layers.13.self_attn.k_proj - target: model.layers.13.self_attn.o_proj weight_key: base_model.model.model.layers.13.self_attn.o_proj - target: model.layers.13.self_attn.q_proj weight_key: base_model.model.model.layers.13.self_attn.q_proj - target: model.layers.13.self_attn.v_proj weight_key: base_model.model.model.layers.13.self_attn.v_proj - target: model.layers.14.mlp.down_proj weight_key: base_model.model.model.layers.14.mlp.down_proj - target: model.layers.14.mlp.gate_proj weight_key: base_model.model.model.layers.14.mlp.gate_proj - target: model.layers.14.mlp.up_proj weight_key: base_model.model.model.layers.14.mlp.up_proj - target: model.layers.14.self_attn.k_proj weight_key: base_model.model.model.layers.14.self_attn.k_proj - target: model.layers.14.self_attn.o_proj weight_key: base_model.model.model.layers.14.self_attn.o_proj - target: model.layers.14.self_attn.q_proj weight_key: base_model.model.model.layers.14.self_attn.q_proj - target: model.layers.14.self_attn.v_proj weight_key: base_model.model.model.layers.14.self_attn.v_proj - target: model.layers.15.mlp.down_proj weight_key: base_model.model.model.layers.15.mlp.down_proj - target: model.layers.15.mlp.gate_proj weight_key: base_model.model.model.layers.15.mlp.gate_proj - target: model.layers.15.mlp.up_proj weight_key: base_model.model.model.layers.15.mlp.up_proj - target: model.layers.15.self_attn.k_proj weight_key: base_model.model.model.layers.15.self_attn.k_proj - target: model.layers.15.self_attn.o_proj weight_key: base_model.model.model.layers.15.self_attn.o_proj - target: model.layers.15.self_attn.q_proj weight_key: base_model.model.model.layers.15.self_attn.q_proj - target: model.layers.15.self_attn.v_proj weight_key: base_model.model.model.layers.15.self_attn.v_proj - target: model.layers.16.mlp.down_proj weight_key: base_model.model.model.layers.16.mlp.down_proj - target: model.layers.16.mlp.gate_proj weight_key: base_model.model.model.layers.16.mlp.gate_proj - target: model.layers.16.mlp.up_proj weight_key: base_model.model.model.layers.16.mlp.up_proj - target: model.layers.16.self_attn.k_proj weight_key: base_model.model.model.layers.16.self_attn.k_proj - target: model.layers.16.self_attn.o_proj weight_key: base_model.model.model.layers.16.self_attn.o_proj - target: model.layers.16.self_attn.q_proj weight_key: base_model.model.model.layers.16.self_attn.q_proj - target: model.layers.16.self_attn.v_proj weight_key: base_model.model.model.layers.16.self_attn.v_proj - target: model.layers.17.mlp.down_proj weight_key: base_model.model.model.layers.17.mlp.down_proj - target: model.layers.17.mlp.gate_proj weight_key: base_model.model.model.layers.17.mlp.gate_proj - target: model.layers.17.mlp.up_proj weight_key: base_model.model.model.layers.17.mlp.up_proj - target: model.layers.17.self_attn.k_proj weight_key: base_model.model.model.layers.17.self_attn.k_proj - target: model.layers.17.self_attn.o_proj weight_key: base_model.model.model.layers.17.self_attn.o_proj - target: model.layers.17.self_attn.q_proj weight_key: base_model.model.model.layers.17.self_attn.q_proj - target: model.layers.17.self_attn.v_proj weight_key: base_model.model.model.layers.17.self_attn.v_proj - target: model.layers.18.mlp.down_proj weight_key: base_model.model.model.layers.18.mlp.down_proj - target: model.layers.18.mlp.gate_proj weight_key: base_model.model.model.layers.18.mlp.gate_proj - target: model.layers.18.mlp.up_proj weight_key: base_model.model.model.layers.18.mlp.up_proj - target: model.layers.18.self_attn.k_proj weight_key: base_model.model.model.layers.18.self_attn.k_proj - target: model.layers.18.self_attn.o_proj weight_key: base_model.model.model.layers.18.self_attn.o_proj - target: model.layers.18.self_attn.q_proj weight_key: base_model.model.model.layers.18.self_attn.q_proj - target: model.layers.18.self_attn.v_proj weight_key: base_model.model.model.layers.18.self_attn.v_proj - target: model.layers.19.mlp.down_proj weight_key: base_model.model.model.layers.19.mlp.down_proj - target: model.layers.19.mlp.gate_proj weight_key: base_model.model.model.layers.19.mlp.gate_proj - target: model.layers.19.mlp.up_proj weight_key: base_model.model.model.layers.19.mlp.up_proj - target: model.layers.19.self_attn.k_proj weight_key: base_model.model.model.layers.19.self_attn.k_proj - target: model.layers.19.self_attn.o_proj weight_key: base_model.model.model.layers.19.self_attn.o_proj - target: model.layers.19.self_attn.q_proj weight_key: base_model.model.model.layers.19.self_attn.q_proj - target: model.layers.19.self_attn.v_proj weight_key: base_model.model.model.layers.19.self_attn.v_proj - target: model.layers.2.mlp.down_proj weight_key: base_model.model.model.layers.2.mlp.down_proj - target: model.layers.2.mlp.gate_proj weight_key: base_model.model.model.layers.2.mlp.gate_proj - target: model.layers.2.mlp.up_proj weight_key: base_model.model.model.layers.2.mlp.up_proj - target: model.layers.2.self_attn.k_proj weight_key: base_model.model.model.layers.2.self_attn.k_proj - target: model.layers.2.self_attn.o_proj weight_key: base_model.model.model.layers.2.self_attn.o_proj - target: model.layers.2.self_attn.q_proj weight_key: base_model.model.model.layers.2.self_attn.q_proj - target: model.layers.2.self_attn.v_proj weight_key: base_model.model.model.layers.2.self_attn.v_proj - target: model.layers.20.mlp.down_proj weight_key: base_model.model.model.layers.20.mlp.down_proj - target: model.layers.20.mlp.gate_proj weight_key: base_model.model.model.layers.20.mlp.gate_proj - target: model.layers.20.mlp.up_proj weight_key: base_model.model.model.layers.20.mlp.up_proj - target: model.layers.20.self_attn.k_proj weight_key: base_model.model.model.layers.20.self_attn.k_proj - target: model.layers.20.self_attn.o_proj weight_key: base_model.model.model.layers.20.self_attn.o_proj - target: model.layers.20.self_attn.q_proj weight_key: base_model.model.model.layers.20.self_attn.q_proj - target: model.layers.20.self_attn.v_proj weight_key: base_model.model.model.layers.20.self_attn.v_proj - target: model.layers.21.mlp.down_proj weight_key: base_model.model.model.layers.21.mlp.down_proj - target: model.layers.21.mlp.gate_proj weight_key: base_model.model.model.layers.21.mlp.gate_proj - target: model.layers.21.mlp.up_proj weight_key: base_model.model.model.layers.21.mlp.up_proj - target: model.layers.21.self_attn.k_proj weight_key: base_model.model.model.layers.21.self_attn.k_proj - target: model.layers.21.self_attn.o_proj weight_key: base_model.model.model.layers.21.self_attn.o_proj - target: model.layers.21.self_attn.q_proj weight_key: base_model.model.model.layers.21.self_attn.q_proj - target: model.layers.21.self_attn.v_proj weight_key: base_model.model.model.layers.21.self_attn.v_proj - target: model.layers.22.mlp.down_proj weight_key: base_model.model.model.layers.22.mlp.down_proj - target: model.layers.22.mlp.gate_proj weight_key: base_model.model.model.layers.22.mlp.gate_proj - target: model.layers.22.mlp.up_proj weight_key: base_model.model.model.layers.22.mlp.up_proj - target: model.layers.22.self_attn.k_proj weight_key: base_model.model.model.layers.22.self_attn.k_proj - target: model.layers.22.self_attn.o_proj weight_key: base_model.model.model.layers.22.self_attn.o_proj - target: model.layers.22.self_attn.q_proj weight_key: base_model.model.model.layers.22.self_attn.q_proj - target: model.layers.22.self_attn.v_proj weight_key: base_model.model.model.layers.22.self_attn.v_proj - target: model.layers.23.mlp.down_proj weight_key: base_model.model.model.layers.23.mlp.down_proj - target: model.layers.23.mlp.gate_proj weight_key: base_model.model.model.layers.23.mlp.gate_proj - target: model.layers.23.mlp.up_proj weight_key: base_model.model.model.layers.23.mlp.up_proj - target: model.layers.23.self_attn.k_proj weight_key: base_model.model.model.layers.23.self_attn.k_proj - target: model.layers.23.self_attn.o_proj weight_key: base_model.model.model.layers.23.self_attn.o_proj - target: model.layers.23.self_attn.q_proj weight_key: base_model.model.model.layers.23.self_attn.q_proj - target: model.layers.23.self_attn.v_proj weight_key: base_model.model.model.layers.23.self_attn.v_proj - target: model.layers.24.mlp.down_proj weight_key: base_model.model.model.layers.24.mlp.down_proj - target: model.layers.24.mlp.gate_proj weight_key: base_model.model.model.layers.24.mlp.gate_proj - target: model.layers.24.mlp.up_proj weight_key: base_model.model.model.layers.24.mlp.up_proj - target: model.layers.24.self_attn.k_proj weight_key: base_model.model.model.layers.24.self_attn.k_proj - target: model.layers.24.self_attn.o_proj weight_key: base_model.model.model.layers.24.self_attn.o_proj - target: model.layers.24.self_attn.q_proj weight_key: base_model.model.model.layers.24.self_attn.q_proj - target: model.layers.24.self_attn.v_proj weight_key: base_model.model.model.layers.24.self_attn.v_proj - target: model.layers.25.mlp.down_proj weight_key: base_model.model.model.layers.25.mlp.down_proj - target: model.layers.25.mlp.gate_proj weight_key: base_model.model.model.layers.25.mlp.gate_proj - target: model.layers.25.mlp.up_proj weight_key: base_model.model.model.layers.25.mlp.up_proj - target: model.layers.25.self_attn.k_proj weight_key: base_model.model.model.layers.25.self_attn.k_proj - target: model.layers.25.self_attn.o_proj weight_key: base_model.model.model.layers.25.self_attn.o_proj - target: model.layers.25.self_attn.q_proj weight_key: base_model.model.model.layers.25.self_attn.q_proj - target: model.layers.25.self_attn.v_proj weight_key: base_model.model.model.layers.25.self_attn.v_proj - target: model.layers.26.mlp.down_proj weight_key: base_model.model.model.layers.26.mlp.down_proj - target: model.layers.26.mlp.gate_proj weight_key: base_model.model.model.layers.26.mlp.gate_proj - target: model.layers.26.mlp.up_proj weight_key: base_model.model.model.layers.26.mlp.up_proj - target: model.layers.26.self_attn.k_proj weight_key: base_model.model.model.layers.26.self_attn.k_proj - target: model.layers.26.self_attn.o_proj weight_key: base_model.model.model.layers.26.self_attn.o_proj - target: model.layers.26.self_attn.q_proj weight_key: base_model.model.model.layers.26.self_attn.q_proj - target: model.layers.26.self_attn.v_proj weight_key: base_model.model.model.layers.26.self_attn.v_proj - target: model.layers.27.mlp.down_proj weight_key: base_model.model.model.layers.27.mlp.down_proj - target: model.layers.27.mlp.gate_proj weight_key: base_model.model.model.layers.27.mlp.gate_proj - target: model.layers.27.mlp.up_proj weight_key: base_model.model.model.layers.27.mlp.up_proj - target: model.layers.27.self_attn.k_proj weight_key: base_model.model.model.layers.27.self_attn.k_proj - target: model.layers.27.self_attn.o_proj weight_key: base_model.model.model.layers.27.self_attn.o_proj - target: model.layers.27.self_attn.q_proj weight_key: base_model.model.model.layers.27.self_attn.q_proj - target: model.layers.27.self_attn.v_proj weight_key: base_model.model.model.layers.27.self_attn.v_proj - target: model.layers.3.mlp.down_proj weight_key: base_model.model.model.layers.3.mlp.down_proj - target: model.layers.3.mlp.gate_proj weight_key: base_model.model.model.layers.3.mlp.gate_proj - target: model.layers.3.mlp.up_proj weight_key: base_model.model.model.layers.3.mlp.up_proj - target: model.layers.3.self_attn.k_proj weight_key: base_model.model.model.layers.3.self_attn.k_proj - target: model.layers.3.self_attn.o_proj weight_key: base_model.model.model.layers.3.self_attn.o_proj - target: model.layers.3.self_attn.q_proj weight_key: base_model.model.model.layers.3.self_attn.q_proj - target: model.layers.3.self_attn.v_proj weight_key: base_model.model.model.layers.3.self_attn.v_proj - target: model.layers.4.mlp.down_proj weight_key: base_model.model.model.layers.4.mlp.down_proj - target: model.layers.4.mlp.gate_proj weight_key: base_model.model.model.layers.4.mlp.gate_proj - target: model.layers.4.mlp.up_proj weight_key: base_model.model.model.layers.4.mlp.up_proj - target: model.layers.4.self_attn.k_proj weight_key: base_model.model.model.layers.4.self_attn.k_proj - target: model.layers.4.self_attn.o_proj weight_key: base_model.model.model.layers.4.self_attn.o_proj - target: model.layers.4.self_attn.q_proj weight_key: base_model.model.model.layers.4.self_attn.q_proj - target: model.layers.4.self_attn.v_proj weight_key: base_model.model.model.layers.4.self_attn.v_proj - target: model.layers.5.mlp.down_proj weight_key: base_model.model.model.layers.5.mlp.down_proj - target: model.layers.5.mlp.gate_proj weight_key: base_model.model.model.layers.5.mlp.gate_proj - target: model.layers.5.mlp.up_proj weight_key: base_model.model.model.layers.5.mlp.up_proj - target: model.layers.5.self_attn.k_proj weight_key: base_model.model.model.layers.5.self_attn.k_proj - target: model.layers.5.self_attn.o_proj weight_key: base_model.model.model.layers.5.self_attn.o_proj - target: model.layers.5.self_attn.q_proj weight_key: base_model.model.model.layers.5.self_attn.q_proj - target: model.layers.5.self_attn.v_proj weight_key: base_model.model.model.layers.5.self_attn.v_proj - target: model.layers.6.mlp.down_proj weight_key: base_model.model.model.layers.6.mlp.down_proj - target: model.layers.6.mlp.gate_proj weight_key: base_model.model.model.layers.6.mlp.gate_proj - target: model.layers.6.mlp.up_proj weight_key: base_model.model.model.layers.6.mlp.up_proj - target: model.layers.6.self_attn.k_proj weight_key: base_model.model.model.layers.6.self_attn.k_proj - target: model.layers.6.self_attn.o_proj weight_key: base_model.model.model.layers.6.self_attn.o_proj - target: model.layers.6.self_attn.q_proj weight_key: base_model.model.model.layers.6.self_attn.q_proj - target: model.layers.6.self_attn.v_proj weight_key: base_model.model.model.layers.6.self_attn.v_proj - target: model.layers.7.mlp.down_proj weight_key: base_model.model.model.layers.7.mlp.down_proj - target: model.layers.7.mlp.gate_proj weight_key: base_model.model.model.layers.7.mlp.gate_proj - target: model.layers.7.mlp.up_proj weight_key: base_model.model.model.layers.7.mlp.up_proj - target: model.layers.7.self_attn.k_proj weight_key: base_model.model.model.layers.7.self_attn.k_proj - target: model.layers.7.self_attn.o_proj weight_key: base_model.model.model.layers.7.self_attn.o_proj - target: model.layers.7.self_attn.q_proj weight_key: base_model.model.model.layers.7.self_attn.q_proj - target: model.layers.7.self_attn.v_proj weight_key: base_model.model.model.layers.7.self_attn.v_proj - target: model.layers.8.mlp.down_proj weight_key: base_model.model.model.layers.8.mlp.down_proj - target: model.layers.8.mlp.gate_proj weight_key: base_model.model.model.layers.8.mlp.gate_proj - target: model.layers.8.mlp.up_proj weight_key: base_model.model.model.layers.8.mlp.up_proj - target: model.layers.8.self_attn.k_proj weight_key: base_model.model.model.layers.8.self_attn.k_proj - target: model.layers.8.self_attn.o_proj weight_key: base_model.model.model.layers.8.self_attn.o_proj - target: model.layers.8.self_attn.q_proj weight_key: base_model.model.model.layers.8.self_attn.q_proj - target: model.layers.8.self_attn.v_proj weight_key: base_model.model.model.layers.8.self_attn.v_proj - target: model.layers.9.mlp.down_proj weight_key: base_model.model.model.layers.9.mlp.down_proj - target: model.layers.9.mlp.gate_proj weight_key: base_model.model.model.layers.9.mlp.gate_proj - target: model.layers.9.mlp.up_proj weight_key: base_model.model.model.layers.9.mlp.up_proj - target: model.layers.9.self_attn.k_proj weight_key: base_model.model.model.layers.9.self_attn.k_proj - target: model.layers.9.self_attn.o_proj weight_key: base_model.model.model.layers.9.self_attn.o_proj - target: model.layers.9.self_attn.q_proj weight_key: base_model.model.model.layers.9.self_attn.q_proj - target: model.layers.9.self_attn.v_proj weight_key: base_model.model.model.layers.9.self_attn.v_proj package: tokenizer: special_tokens: eos_token_id: - 151645