| { |
| "analysis_sha256": "d688d7f271f8bd5c850515e3fc5e01138710005fbdee7719a2beb4895c993a8d", |
| "architecture_profile": { |
| "adapter_id": "mistral3-dense-v1", |
| "audio_present": false, |
| "config_model_type": "ministral3", |
| "dense": true, |
| "mtp_declared": false, |
| "notes": [ |
| "Mistral3 multimodal shells (including Ministral-3): language path is optimized; vision tower is stripped by MLX-LM sanitize and preserved only when present as protected tensors in the source inventory.", |
| "Nemotron-3-Embed (model_type=ministral3) is feature-extraction only; do not claim generative or MTP metrics." |
| ], |
| "optimization_scope": "text-path", |
| "product_family": "mistral3", |
| "support_level": "supported", |
| "support_tier": "convertible", |
| "text_layer_count": 16, |
| "vision_present": false |
| }, |
| "assignments": [ |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "embed_tokens", |
| "outlier_strategy": "none", |
| "parameters": 268435456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule embedding-8: Embedding floor 8-bit affine", |
| "role": "embedding", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "embed_tokens.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.0.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.0.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.0.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.0.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.1.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.1.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.1.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.1.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.10.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.10.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.10.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.10.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.11.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.11.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.11.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.11.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.12.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.12.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.12.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.12.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.13.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.13.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.13.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.13.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.14.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.14.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.14.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.14.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.15.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.15.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.15.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.15.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.2.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.2.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.2.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.2.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.3.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.3.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.3.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.3.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.4.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.4.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.4.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.4.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.5.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.5.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.5.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.5.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.6.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.6.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.6.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.6.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.7.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.7.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.7.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.7.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.8.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.8.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.8.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.8.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.9.input_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 12582912, |
| "predicted_loss": 0.0, |
| "reason": "manual rule mlp-mxfp8: AXQ-mxfp8 MLP trunk", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "layers.9.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 6291456, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 8, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "layers.9.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 2097152, |
| "predicted_loss": 0.0, |
| "reason": "manual rule attention-mxfp8: AXQ-mxfp8 attention trunk", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 9.0 |
| }, |
| "tensor": "layers.9.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "norm", |
| "outlier_strategy": "none", |
| "parameters": 2048, |
| "predicted_loss": 0.0, |
| "reason": "manual rule protect-norms: Protection floor — norms stay BF16", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "norm.weight" |
| } |
| ], |
| "calibration": null, |
| "candidate_bits": [ |
| 8, |
| 16 |
| ], |
| "candidate_group_sizes": [], |
| "constraints": { |
| "effective_bpw_limit": 16.0, |
| "lm_head_min_bits": 16, |
| "max_model_size_ratio_to_uniform4": 1.1, |
| "minimum_mtp_acceptance_retention": 0.95, |
| "minimum_mtp_speedup": 1.2, |
| "minimum_quality_retention": 0.98 |
| }, |
| "cost_model": "abstract-bpw", |
| "created_at": "2026-10-04T02:08:58.748238Z", |
| "effective_bpw": 9.000414655467978, |
| "evidence_kind": "architecture_prior", |
| "global_validation_required": true, |
| "group_size": 32, |
| "hardware": { |
| "name": "ax-engine-apple-silicon-affine-dwq-v3", |
| "runtime": "ax-engine", |
| "supported_bits": [ |
| 2, |
| 3, |
| 4, |
| 6, |
| 8, |
| 16 |
| ], |
| "supported_group_sizes": [ |
| 32, |
| 64, |
| 128 |
| ], |
| "supported_methods": [ |
| "affine", |
| "awq", |
| "dwq", |
| "gptq", |
| "gptq-act", |
| "bf16" |
| ] |
| }, |
| "kernel_latency_host_id": null, |
| "kernel_latency_sha256": null, |
| "kv_cache": null, |
| "method_near_ties": [], |
| "method_near_ties_omitted": 0, |
| "mtp": { |
| "candidate_bits": [ |
| 8, |
| 16 |
| ], |
| "min_bits": 8, |
| "mode": "protected", |
| "optimize_for_acceptance": true, |
| "preserve_external_sidecar": true, |
| "protect_norms": true, |
| "protect_output_head": true |
| }, |
| "mtp_distribution": {}, |
| "nominal_bpw": 8.000473891963402, |
| "objective": { |
| "cosine_distance": 0.03, |
| "decode_latency_cost": 0.07, |
| "hidden_state_error": 0.1, |
| "long_context_loss": 0.05, |
| "mtp_acceptance_loss": 0.22, |
| "output_kl": 0.15, |
| "peak_memory_cost": 0.03, |
| "prefill_latency_cost": 0.03, |
| "task_loss_delta": 0.2, |
| "token_disagreement": 0.12 |
| }, |
| "primary_runtime": "ax-engine", |
| "profile": "agent-coding", |
| "quantizer": "axquant", |
| "random_seed": 20260728, |
| "schema_version": "axquant.plan.v2", |
| "software_versions": { |
| "ax_engine": "7.5.7", |
| "axquant": "1.9.0", |
| "mlx": "0.32.1", |
| "mlx_lm": "0.31.3", |
| "pydantic": "2.13.4", |
| "python": "3.12.13", |
| "safetensors": "0.8.0" |
| }, |
| "source_model": { |
| "architecture": "Ministral3Model", |
| "format": "mlx", |
| "local_path": null, |
| "model_id": "nvidia/Nemotron-3-Embed-1B-BF16", |
| "revision": "c0c9fea93ea424587517f2c59e20db9f1d6bf615" |
| }, |
| "status": "planned", |
| "target_bpw": 16.0, |
| "target_class": "16p0bpw", |
| "target_mode": "low-memory", |
| "warnings": [ |
| "Manual assignments are unmeasured development evidence.", |
| "Conversion requires --allow-unmeasured and cannot pass publication gates." |
| ], |
| "weight_distribution": { |
| "8bit": { |
| "fraction": 0.9999407635045747, |
| "parameters": 1140850688 |
| }, |
| "bf16": { |
| "fraction": 5.923649542532701e-05, |
| "parameters": 67584 |
| } |
| } |
| } |
|
|
