diff --git "a/sglang_patch/nightly-dev-20260908-20ca564b/sglang/srt/models/qwen3_5.py" "b/sglang_patch/nightly-dev-20260908-20ca564b/sglang/srt/models/qwen3_5.py" new file mode 100644--- /dev/null +++ "b/sglang_patch/nightly-dev-20260908-20ca564b/sglang/srt/models/qwen3_5.py" @@ -0,0 +1,2770 @@ +# Copyright 2025 Qwen Team +# Copyright 2025 SGLang Team +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Inference-only Qwen3.5 model and Qwen3.5 MoE model compatible with HuggingFace weights.""" + +import logging +import os +from functools import lru_cache +from typing import Iterable, Optional, Set, Tuple, Union + +import torch +import torch.nn as nn +import triton + +# Layers - Attention +from sglang.kernels.ops.attention.fla.layernorm_gated import RMSNorm as RMSNormGated +from sglang.kernels.ops.attention.triton_gdn_fused_proj import ( + fused_qkvzba_split_reshape_cat_contiguous, + qwen3_5_gdn_prefill_projection_views, +) +from sglang.kernels.ops.elementwise.elementwise import fused_sigmoid_mul + +# Configs +from sglang.srt.configs.qwen3_5 import ( + Qwen3_5Config, + Qwen3_5MoeConfig, + Qwen3_5TextConfig, +) + +# Distributed +from sglang.srt.distributed import get_pp_group +from sglang.srt.environ import envs +from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder +from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation +from sglang.srt.layers.attention.mamba.mamba import mamba_v2_sharded_weight_loader +from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes +from sglang.srt.layers.dp_attention import ( + is_dp_attention_enabled, +) + +# Layers - Others +from sglang.srt.layers.layernorm import GemmaRMSNorm + +# Layers - Linear +from sglang.srt.layers.linear import ( + ColumnParallelLinear, + MergedColumnParallelLinear, + QKVParallelLinear, + RowParallelLinear, +) +from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE +from sglang.srt.layers.moe.utils import ( + is_shared_experts_fusion_disabled, +) +from sglang.srt.layers.parameter import ( + BlockQuantScaleParameter, + PerTensorScaleParameter, +) +from sglang.srt.layers.quantization.base_config import QuantizationConfig +from sglang.srt.layers.radix_attention import RadixAttention +from sglang.srt.layers.radix_linear_attention import RadixLinearAttention +from sglang.srt.layers.rotary_embedding import get_rope +from sglang.srt.layers.utils import PPMissingLayer, get_layer_id +from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding +from sglang.srt.model_executor.cuda_graph_config import ( + Backend, + Phase, + check_cuda_graph_backend, +) +from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors +from sglang.srt.model_executor.runner import get_is_capture_mode +from sglang.srt.model_loader.weight_utils import ( + default_weight_loader, + sharded_weight_loader, +) +from sglang.srt.models.qwen2_moe import ( + Qwen2MoeMLP, + Qwen2MoeSparseMoeBlock, + can_fuse_shared_expert, +) + +# Models +from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration +from sglang.srt.models.utils import ( + WeightsMapper, + fused_qk_gemma_rmsnorm, + fused_qk_gemma_rmsnorm_with_gate, +) +from sglang.srt.runtime_context import ( + get_exec, + get_forward, + get_parallel, + get_stream, +) + +# Utils +from sglang.srt.utils import ( + LazyValue, + add_prefix, + cpu_has_amx_support, + get_bool_env_var, + is_cpu, + is_cuda, + is_gfx95_supported, + is_hip, + is_npu, + is_xpu, + make_layers, + set_weight_attrs, + use_intel_amx_backend, +) +from sglang.srt.utils.hf_transformers_utils import get_processor, get_rope_config + +logger = logging.getLogger(__name__) +_is_cuda = is_cuda() +_is_npu = is_npu() +_is_cpu = is_cpu() +_is_gfx95 = is_gfx95_supported() +_is_hip = is_hip() +_is_xpu = is_xpu() +_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip +_hip_use_alt_stream = get_bool_env_var("SGLANG_ALT_STREAM") and _is_hip +_gdn_use_alt_stream = _is_cuda or ( + get_bool_env_var("SGLANG_GDN_QKVZ_BA_ALT_STREAM", "False") and _hip_use_alt_stream +) +_qknorm_use_alt_stream = _is_cuda or ( + get_bool_env_var("SGLANG_QK_NORM_ALT_STREAM", "False") and _hip_use_alt_stream +) +_gdn_decode_fused_proj_conv = ( + _is_cuda and envs.SGLANG_ENABLE_GDN_DECODE_FUSED_PROJ_CONV.get() +) +_is_amx_available = cpu_has_amx_support() +_is_xpu = is_xpu() + +# Head-group ratios (num_v_heads // num_k_heads) served by the fused +# split/reshape/cat Triton kernel. On AMD/aiter the ratio-8 layout is also +# covered by the fused kernel, which removes the two `.contiguous()` copies +# plus the `torch.cat` of the unfused fallback. On CUDA the ratio-3 dense 27B +# layout is handled by the Triton kernel's per-head walk (the CPU fused op +# still requires a power-of-two group). Other backends keep the original +# tuple so their control flow is unchanged. +_GDN_FUSED_QKVZBA_RATIOS = ( + (1, 2, 4, 8) if _use_aiter else (1, 2, 3, 4) if _is_cuda else (1, 2, 4) +) + +cached_get_processor = lru_cache(get_processor) + + +def _disable_shared_experts_fusion() -> bool: + # Resolved lazily: the flag is written by the owning model's gate before + # its layers build (per runner); models without a gate see the config + # intent through the accessor's fallback. + # The deferred-finalize ABI needs the shared expert as a separate, gated + # local contribution; it cannot consume a shared slot fused into routed MoE. + return bool( + envs.SGLANG_FLASHINFER_MNNVL_CUTEDSL_AR_FUSION.get() + or is_shared_experts_fusion_disabled() + ) + + +def _maybe_enable_silu_fp4_quant_fusion(mlp: nn.Module) -> None: + """Fuse SiLU+mul with the down_proj NVFP4 input quantization. + + Replaces the separate act_and_mul and per-token FP4 quantize kernels with + one FlashInfer kernel and feeds down_proj a prequantized (fp4, scale) + tuple. Enabled when both dense-MLP projections use the NVFP4 (W4A4) + linear method; kill switch: SGLANG_DISABLE_SILU_FP4_QUANT_FUSION=1. + """ + if os.environ.get("SGLANG_DISABLE_SILU_FP4_QUANT_FUSION", "0") == "1": + return + from sglang.srt.layers.quantization.modelopt_quant import ModelOptFp4LinearMethod + + if not ( + isinstance(mlp.gate_up_proj.quant_method, ModelOptFp4LinearMethod) + and mlp.gate_up_proj.quant_method.quant_mode == "w4a4" + and isinstance(mlp.down_proj.quant_method, ModelOptFp4LinearMethod) + ): + return + try: + from flashinfer import silu_and_mul_scaled_nvfp4_experts_quantize # noqa: F401 + except ImportError: + return + mlp._enable_silu_fp4_quant_fusion = True + mlp.down_proj._accepts_prequantized_fp4 = True + logger.info("Enabled fused SiLU+mul+FP4-quant for dense MLP down_proj input.") + + +def _use_mnnvl_cutedsl_fusion(config: Qwen3_5TextConfig, is_nextn: bool) -> bool: + return bool( + not is_nextn + and config.model_type == "qwen3_5_moe_text" + and envs.SGLANG_FLASHINFER_MNNVL_CUTEDSL_AR_FUSION.get() + ) + + +def _layer_communicator_class(config: Qwen3_5TextConfig, is_nextn: bool): + if _use_mnnvl_cutedsl_fusion(config, is_nextn): + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + Qwen35FlashInferLayerCommunicator, + ) + + return Qwen35FlashInferLayerCommunicator + return LayerCommunicator + + +if _is_cuda: + from sglang.kernels.ops.attention.fused_qk_rmsnorm_rope_gate import ( + fused_qk_gemma_rmsnorm_rope_gate, + ) + +if _is_cpu: + _fused_sigmoid_mul_cpu = torch.ops.sgl_kernel.fused_sigmoid_mul_cpu + + def fused_sigmoid_mul(x, gate, inplace=True): + if not inplace: + x = x.clone() + _fused_sigmoid_mul_cpu(x, gate) + return x + + fused_qk_gemma_rmsnorm = torch.ops.sgl_kernel.fused_qk_gemma_rmsnorm_cpu + fused_qk_gemma_rmsnorm_with_gate = ( + torch.ops.sgl_kernel.fused_qk_gemma_rmsnorm_with_gate_cpu + ) + fused_qkvzba_split_reshape_cat_contiguous = ( + torch.ops.sgl_kernel.fused_qkvzba_split_reshape_cat_contiguous_cpu + ) + + +@lru_cache(maxsize=1) +def _enable_qwen35_fused_ar_quant() -> bool: + """Gate the fused AR+RMSNorm+per-group-FP8-quant path for Qwen3.5. + + The single-kernel backend is ROCm/aiter/gfx95-only. The model gate stays + tied to ROCm/aiter so non-gfx95 HIP can keep the existing 2-kernel fallback + behavior for tuple handoff when this branch is used. It replaces the + existing ``--enable-aiter-allreduce-fusion`` 3-kernel path + (AR → RMSNorm → per-group quant) with either a single fused kernel (when + the fully-fused variant is eligible) or a 2-kernel path + (fused AR+RMSNorm + separate per-group quant) that still saves one + kernel launch vs. baseline. The LayerCommunicator gracefully falls back + to ``forward_with_allreduce_fusion`` (plain AR+RMSNorm) when the fused + quant helper returns ``None``, so turning this on never regresses the + AR+RMSNorm fusion itself. + + Opt-out: set ``SGLANG_DISABLE_FUSED_AR_QUANT=1`` to fall back to the + unmodified AR+RMSNorm fusion path. + """ + if not _use_aiter: + return False + if get_bool_env_var("SGLANG_DISABLE_FUSED_AR_QUANT", default="false"): + return False + return bool(get_exec().comm.enable_aiter_allreduce_fusion) + + +def _linear_accepts_fp8_tuple(linear: nn.Module) -> bool: + quant_method = getattr(linear, "quant_method", None) + return quant_method.__class__.__name__ == "Fp8LinearMethod" and ( + getattr(quant_method, "block_quant", False) + or getattr(quant_method, "use_mxfp8", False) + ) + + +def _select_fused_ar_input_for_linear(hidden_states, linear: nn.Module): + if not isinstance(hidden_states, tuple): + return hidden_states + if len(hidden_states) == 3: + hs_bf16, hs_fp8, hs_scale = hidden_states + if _linear_accepts_fp8_tuple(linear): + return (hs_fp8, hs_scale) + return hs_bf16 + if len(hidden_states) == 2 and _linear_accepts_fp8_tuple(linear): + return hidden_states + raise TypeError( + f"{linear.__class__.__name__} cannot consume fused AR quant tuple input" + ) + + +def _finish_mlp_output(hidden_states, *, expect_deferred: bool): + if not expect_deferred: + if not isinstance(hidden_states, torch.Tensor): + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + Qwen35MoeFinalizeHandoff, + ) + + if isinstance(hidden_states, Qwen35MoeFinalizeHandoff): + raise RuntimeError("unexpected deferred-finalize handoff") + hidden_states._sglang_needs_allreduce_fusion = True + return hidden_states + + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + Qwen35MoeFinalizeHandoff, + ) + + if not isinstance(hidden_states, Qwen35MoeFinalizeHandoff): + raise RuntimeError("Qwen3.5 expected a FlashInfer deferred-finalize handoff") + return hidden_states + + +if _is_npu: + from sgl_kernel_npu.norm.split_qkv_rmsnorm_rope import ( + split_qkvgate_gemma_rmsnorm_rope, + ) + + +class Qwen3_5GatedDeltaNet(nn.Module): + def __init__( + self, + config: Qwen3_5TextConfig, + layer_id: int, + quant_config: Optional[QuantizationConfig] = None, + alt_stream: Optional[torch.cuda.Stream] = None, + prefix: str = "", + ) -> None: + super().__init__() + self.config = config + self.attn_tp_rank = get_parallel().attn_tp_rank + self.attn_tp_size = get_parallel().attn_tp_size + self.hidden_size = config.hidden_size + self.num_v_heads = ( + config.linear_num_value_heads + if not _is_cpu + else config.linear_num_value_heads_cpu + ) + self.num_k_heads = ( + config.linear_num_key_heads + if not _is_cpu + else config.linear_num_key_heads_cpu + ) + self.head_k_dim = config.linear_key_head_dim + self.head_v_dim = config.linear_value_head_dim + self.key_dim = self.head_k_dim * self.num_k_heads + self.value_dim = self.head_v_dim * self.num_v_heads + self.alt_stream = alt_stream + + self.conv_kernel_size = config.linear_conv_kernel_dim + self.layer_id = layer_id + self.activation = config.hidden_act + self.output_gate_type = config.output_gate_type + self.layer_norm_epsilon = config.rms_norm_eps + + # Conv1d layer + self.conv_dim = self.key_dim * 2 + self.value_dim + self.conv1d = ColumnParallelLinear( + input_size=self.conv_kernel_size, + output_size=self.conv_dim, + bias=False, + quant_config=None, + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + prefix=add_prefix("conv1d", prefix), + ) + self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1) + + # projection of the input hidden states + self.in_proj_qkvz = self.create_qkvz_proj( + hidden_size=self.hidden_size, + key_dim=self.key_dim, + value_dim=self.value_dim, + quant_config=quant_config, + prefix=add_prefix("in_proj_qkvz", prefix), + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + ) + + self.in_proj_ba = self.create_ba_proj( + hidden_size=self.hidden_size, + num_v_heads=self.num_v_heads, + quant_config=quant_config, + prefix=add_prefix("in_proj_ba", prefix), + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + ) + + # Override weight loaders for packed checkpoint format. + # Important: for FP8, this must cover not only `.weight` but also + # `weight_scale_inv` / `weight_scale` / `input_scale` if present. + self._bind_packed_weight_loaders(self.in_proj_qkvz) + self._bind_packed_weight_loaders(self.in_proj_ba) + self._fused_input_proj_cpu_enabled = LazyValue( + lambda: ( + _is_cpu + and self.in_proj_qkvz._parameters.get("weight") is not None + and self.in_proj_ba._parameters.get("weight") is not None + and self.in_proj_qkvz._parameters["weight"].dtype == torch.bfloat16 + and self.in_proj_ba._parameters["weight"].dtype == torch.bfloat16 + and self.in_proj_qkvz.bias is None + and self.in_proj_ba.bias is None + and use_intel_amx_backend(self.in_proj_qkvz) + and use_intel_amx_backend(self.in_proj_ba) + and ( + self.in_proj_qkvz.weight.size(0) % 32 == 0 + and self.in_proj_ba.weight.size(0) % 32 == 0 + ) + ) + ) + + # Conv1d weight loader setup + query_key_settings = (self.key_dim, 0, False) + value_settings = (self.value_dim, 0, False) + + self._override_weight_loader( + self.conv1d.weight, + mamba_v2_sharded_weight_loader( + [ + query_key_settings, + query_key_settings, + value_settings, + ], + self.attn_tp_size, + self.attn_tp_rank, + ), + ) + + # State parameters + self.dt_bias = nn.Parameter( + torch.ones(self.num_v_heads // self.attn_tp_size), + ) + self.A_log = nn.Parameter( + torch.empty(self.num_v_heads // self.attn_tp_size, dtype=torch.float32), + ) + + set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)}) + set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)}) + + conv_weights = self.conv1d.weight.view( + self.conv1d.weight.size(0), self.conv1d.weight.size(2) + ) + self.attn = RadixLinearAttention( + layer_id=layer_id, + num_q_heads=self.num_k_heads // self.attn_tp_size, + num_k_heads=self.num_k_heads // self.attn_tp_size, + num_v_heads=self.num_v_heads // self.attn_tp_size, + head_q_dim=self.head_k_dim, + head_k_dim=self.head_k_dim, + head_v_dim=self.head_v_dim, + conv_weights=conv_weights, + bias=self.conv1d.bias, + activation=self.activation, + A_log=self.A_log, + dt_bias=self.dt_bias, + ) + + self.norm = RMSNormGated( + self.head_v_dim, + eps=self.layer_norm_epsilon, + group_size=None, + norm_before_gate=True, + device=torch.get_device_module().current_device(), + dtype=torch.get_default_dtype(), + **( + {"activation": self.output_gate_type} + if self.output_gate_type is not None + else {} + ), + ) + + self.out_proj = RowParallelLinear( + self.value_dim, + self.hidden_size, + bias=False, + input_is_parallel=True, + reduce_results=False, + quant_config=quant_config, + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + prefix=add_prefix("out_proj", prefix), + ) + + @staticmethod + def _override_weight_loader(param, loader): + """Robustly override loader for: + 1) BasevLLMParameter subclasses: real storage is `_weight_loader` + 2) regular Parameters that already have mutable `weight_loader` + 3) regular Parameters without `weight_loader` yet + """ + if hasattr(param, "_weight_loader"): + # FP8 / quantized BasevLLMParameter path + param._weight_loader = loader + return + + if hasattr(param, "weight_loader"): + # Regular parameter/tensor that already has a mutable attr. + # Do NOT call set_weight_attrs here, because it asserts when + # overwriting an existing attribute. + param.weight_loader = loader + return + + # Fresh attribute on a normal tensor/Parameter + set_weight_attrs(param, {"weight_loader": loader}) + + def _bind_packed_weight_loaders(self, module): + """Bind packed-checkpoint-aware loaders to all relevant params of a merged module.""" + for attr_name in ("weight", "weight_scale_inv", "weight_scale", "input_scale"): + param = getattr(module, attr_name, None) + if param is None: + continue + original_loader = getattr(param, "weight_loader", None) + if original_loader is None: + continue + wrapped_loader = self._make_packed_weight_loader(module, original_loader) + self._override_weight_loader(param, wrapped_loader) + + @staticmethod + def _get_split_sizes_for_param(module, param, loaded_shard_id): + """Return checkpoint-side split sizes for this param type.""" + if isinstance(param, BlockQuantScaleParameter): + # Split by output blocks, not raw output sizes. + block_n, _ = module.quant_method.quant_config.weight_block_size + block_n = 1 if getattr(param, "format_ue8m0", False) else block_n + return [ + (module.output_sizes[idx] + block_n - 1) // block_n + for idx in loaded_shard_id + ] + + if isinstance(param, PerTensorScaleParameter): + # One logical scale per logical shard. + return [1 for _ in loaded_shard_id] + + # Normal weight / non-block quant tensor + return [module.output_sizes[idx] for idx in loaded_shard_id] + + @classmethod + def _make_packed_weight_loader(cls, module, original_weight_loader): + """Wrap the param's original loader so split checkpoints: + - in_proj_qkv + in_proj_z -> merged in_proj_qkvz + - in_proj_b + in_proj_a -> merged in_proj_ba + can load correctly for both normal and FP8 params. + """ + + def weight_loader(param, loaded_weight, loaded_shard_id=None): + # Only intercept split-checkpoint tuple shards. + # int shard_id and None should preserve original behavior. + if isinstance(loaded_shard_id, tuple): + split_sizes = cls._get_split_sizes_for_param( + module, param, loaded_shard_id + ) + + if loaded_weight.numel() == 1: + # Single-element tensor (scalar or [1]): + # broadcast to each logical shard. + chunks = [loaded_weight.view(-1)] * len(loaded_shard_id) + else: + split_dim = getattr(param, "output_dim", 0) + if _is_cpu: + cpu_split_sizes = [] + split_size_sum = sum(split_sizes) + target_size_sim = loaded_weight.size(split_dim) + for i in range(len(split_sizes)): + cpu_split_sizes.append( + int(target_size_sim * split_sizes[i] / split_size_sum) + ) + assert sum(cpu_split_sizes) == target_size_sim, ( + f"Padding the loaded weight failed due to sizes are not divisible cleanly from {cpu_split_sizes} to {target_size_sim}" + ) + chunks = loaded_weight.split(cpu_split_sizes, dim=split_dim) + else: + chunks = loaded_weight.split(split_sizes, dim=split_dim) + + assert len(chunks) == len(loaded_shard_id), ( + f"Chunk/shard mismatch: {len(chunks)=}, " + f"{len(loaded_shard_id)=}, {split_sizes=}" + ) + + for idx, chunk in zip(loaded_shard_id, chunks): + # Delegate each chunk to the param's original int-shard loader. + original_weight_loader(param, chunk, idx) + return + + return original_weight_loader(param, loaded_weight, loaded_shard_id) + + return weight_loader + + def create_qkvz_proj( + self, + hidden_size: int, + key_dim: int, + value_dim: int, + quant_config: QuantizationConfig | None, + prefix: str, + tp_rank: Optional[int] = None, + tp_size: Optional[int] = None, + ) -> MergedColumnParallelLinear: + return MergedColumnParallelLinear( + input_size=hidden_size, + output_sizes=[key_dim, key_dim, value_dim, value_dim], + bias=False, + quant_config=quant_config, + prefix=prefix, + tp_rank=tp_rank, + tp_size=tp_size, + ) + + def create_ba_proj( + self, + hidden_size: int, + num_v_heads: int, + quant_config: QuantizationConfig | None, + prefix: str, + tp_rank: Optional[int] = None, + tp_size: Optional[int] = None, + ) -> MergedColumnParallelLinear: + # Qwen3.5 has separate in_proj_b and in_proj_a weights in the + # checkpoint, which are loaded into the fused in_proj_ba parameter + # via stacked_params_mapping with shard_id 0 and 1 respectively. + return MergedColumnParallelLinear( + input_size=hidden_size, + output_sizes=[num_v_heads, num_v_heads], + bias=False, + quant_config=quant_config, + prefix=prefix, + tp_rank=tp_rank, + tp_size=tp_size, + ) + + def fix_query_key_value_ordering( + self, + mixed_qkvz: torch.Tensor, + mixed_ba: torch.Tensor, + ): + """ + Derives `query`, `key` and `value` tensors from `mixed_qkvzba`. + """ + k_tp = self.key_dim // self.attn_tp_size + v_tp = self.value_dim // self.attn_tp_size + nv_tp = self.num_v_heads // self.attn_tp_size + + # Directly split, no head group reshape + query, key, value, z = mixed_qkvz.split([k_tp, k_tp, v_tp, v_tp], dim=-1) + b, a = mixed_ba.split([nv_tp, nv_tp], dim=-1) + + # value / z reshape to (seq, num_v_heads/tp, head_v_dim) + value = value.reshape(value.size(0), -1, self.head_v_dim) + z = z.reshape(z.size(0), -1, self.head_v_dim) + + return query, key, value, z, b, a + + def _forward_input_proj(self, hidden_states: torch.Tensor): + # AMD/aiter fused AR+RMSNorm+per-group-quant path ships a + # ``(bf16, fp8, scale)`` 3-tuple so the FP8 ``in_proj_qkvz`` can + # consume ``(fp8, scale)`` (skipping its internal quant) while the + # bf16 ``in_proj_ba`` consumes the unquantized bf16. Non-aiter runs skip + # the tuple branch and keep the original control flow below unchanged. + if _use_aiter and isinstance(hidden_states, tuple): + return self._forward_input_proj_fused_quant_amd(hidden_states) + + if ( + _is_cpu + or _is_npu + or check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE) + ): + DUAL_STREAM_TOKEN_THRESHOLD = 0 + else: + DUAL_STREAM_TOKEN_THRESHOLD = 1024 + + seq_len, _ = hidden_states.shape + if ( + self.alt_stream is not None + and get_is_capture_mode() + and seq_len < DUAL_STREAM_TOKEN_THRESHOLD + and _gdn_use_alt_stream + ): + current_stream = torch.cuda.current_stream() + self.alt_stream.wait_stream(current_stream) + projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states) + with torch.cuda.stream(self.alt_stream): + projected_states_ba, _ = self.in_proj_ba(hidden_states) + current_stream.wait_stream(self.alt_stream) + elif self._fused_input_proj_cpu_enabled.value: + projected_states_qkvz, projected_states_ba = ( + torch.ops.sgl_kernel.fused_input_proj_cpu( + hidden_states, + self.in_proj_qkvz.weight, + self.in_proj_ba.weight, + True, + ) + ) + else: + projected_states_qkvz, _ = self.in_proj_qkvz(hidden_states) + projected_states_ba, _ = self.in_proj_ba(hidden_states) + return projected_states_qkvz, projected_states_ba + + def _forward_input_proj_fused_quant_amd(self, hidden_states): + """AMD-only variant for the fused AR+RMSNorm+per-group-quant path. + + ``hidden_states`` is a ``(bf16, fp8, scale)`` 3-tuple produced by the + upstream fused kernel. FP8 ``in_proj_qkvz`` takes ``(fp8, scale)`` + directly; unquantized variants take the bf16 side-output. + """ + hs_bf16 = hidden_states[0] + hs_qkvz = _select_fused_ar_input_for_linear(hidden_states, self.in_proj_qkvz) + seq_len = hs_bf16.shape[0] + + if check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE): + DUAL_STREAM_TOKEN_THRESHOLD = 0 + else: + DUAL_STREAM_TOKEN_THRESHOLD = 1024 + + if ( + self.alt_stream is not None + and get_is_capture_mode() + and seq_len < DUAL_STREAM_TOKEN_THRESHOLD + and _gdn_use_alt_stream + ): + current_stream = torch.cuda.current_stream() + self.alt_stream.wait_stream(current_stream) + projected_states_qkvz, _ = self.in_proj_qkvz(hs_qkvz) + with torch.cuda.stream(self.alt_stream): + projected_states_ba, _ = self.in_proj_ba(hs_bf16) + current_stream.wait_stream(self.alt_stream) + else: + projected_states_qkvz, _ = self.in_proj_qkvz(hs_qkvz) + projected_states_ba, _ = self.in_proj_ba(hs_bf16) + return projected_states_qkvz, projected_states_ba + + def _forward_xpu( + self, + backend: object, + projected_states_qkvz: torch.Tensor, + projected_states_ba: torch.Tensor, + forward_batch: ForwardBatch, + ): + core_attn_out, z = backend.forward_fused_gdn( + self.attn, + forward_batch, + projected_states_qkvz, + projected_states_ba, + ) + + assert core_attn_out is not None, "XPU backend must support fused GDN" + + z_shape_og = z.shape + # reshape input data into 2D tensor + core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1]) + z = z.reshape(-1, z.shape[-1]) + + # Add padding for DP-Attn + if core_attn_out.shape != z.shape: + core_attn_out_pad = torch.zeros_like(z) + core_attn_out_pad[: core_attn_out.shape[0], :] = core_attn_out + core_attn_out = core_attn_out_pad + + core_attn_out = self.norm(core_attn_out, z) + core_attn_out = core_attn_out.reshape(z_shape_og) + core_attn_out = core_attn_out.reshape(*core_attn_out.shape[:-2], -1) + + output, _ = self.out_proj(core_attn_out) + return output + + def forward( + self, + hidden_states: torch.Tensor, + forward_batch: ForwardBatch, + ): + """ + Forward pass with three parts: + 1. Input projection + 2. Core attention (custom op) + 3. Output projection + """ + projected_states_qkvz, projected_states_ba = self._forward_input_proj( + hidden_states + ) + + if _is_xpu and get_exec().mamba.linear_attn_backend == "intel_xpu": + from sglang.srt.model_executor.forward_context import get_attn_backend + + backend = get_attn_backend() + backend = getattr(backend, "linear_attn_backend", backend) + if backend.supports_fused_gdn(self.attn, forward_batch): + return self._forward_xpu( + backend, projected_states_qkvz, projected_states_ba, forward_batch + ) + + use_strided_prefill_z = False + use_fused_decode_proj_conv = ( + _gdn_decode_fused_proj_conv + and forward_batch.forward_mode.is_decode() + and isinstance(projected_states_qkvz, torch.Tensor) + and isinstance(projected_states_ba, torch.Tensor) + ) + use_fused_contiguous_unpack = ( + self.num_v_heads // self.num_k_heads in _GDN_FUSED_QKVZBA_RATIOS + ) + if use_fused_decode_proj_conv: + # GDN owns indexed Conv1D state and the safe unpack/Conv boundary; + # it replaces these temporary B/A placeholders before recurrence. + mixed_qkv = (projected_states_qkvz, projected_states_ba) + z = None + b = projected_states_ba + a = projected_states_ba + elif use_fused_contiguous_unpack and not _is_npu: + if _is_cpu: + num_k_heads_tp = self.num_k_heads // self.attn_tp_size + num_v_heads_tp = self.num_v_heads // self.attn_tp_size + else: + num_k_heads_tp = triton.cdiv(self.num_k_heads, self.attn_tp_size) + num_v_heads_tp = triton.cdiv(self.num_v_heads, self.attn_tp_size) + use_strided_prefill_z = ( + _is_cuda and forward_batch.forward_mode.is_extend_without_speculative() + ) + split_fn = ( + qwen3_5_gdn_prefill_projection_views + if use_strided_prefill_z + else fused_qkvzba_split_reshape_cat_contiguous + ) + mixed_qkv, z, b, a = split_fn( + projected_states_qkvz, + projected_states_ba, + num_k_heads_tp, + num_v_heads_tp, + self.head_k_dim, + self.head_v_dim, + ) + else: + query, key, value, z, b, a = self.fix_query_key_value_ordering( + projected_states_qkvz, projected_states_ba + ) + b = b.contiguous() + a = a.contiguous() + + query, key, value = map( + lambda x: x.reshape(x.shape[0], -1), (query, key, value) + ) + mixed_qkv = torch.cat((query, key, value), dim=-1) + + attn_result = self.attn( + forward_batch, + mixed_qkv=mixed_qkv, + a=a, + b=b, + ) + if use_fused_decode_proj_conv: + if not isinstance(attn_result, tuple) or len(attn_result) != 2: + raise RuntimeError( + "Fused GDN decode projection/Conv1D backend must return " + "(core_attn_out, z)" + ) + core_attn_out, z = attn_result + else: + core_attn_out = attn_result + assert z is not None + + z_shape_og = z.shape + # reshape input data into 2D tensor + core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1]) + if use_strided_prefill_z: + z_flat_shape = (z.numel() // z.shape[-1], z.shape[-1]) + else: + z = z.reshape(-1, z.shape[-1]) + z_flat_shape = z.shape + + # Add padding for DP-Attn + if core_attn_out.shape != z_flat_shape: + core_attn_out_pad = z.new_zeros(z_flat_shape) + core_attn_out_pad[: core_attn_out.shape[0], :] = core_attn_out + core_attn_out = core_attn_out_pad + + core_attn_out = self.norm(core_attn_out, z) + core_attn_out = core_attn_out.reshape(z_shape_og) + core_attn_out = core_attn_out.reshape( + *core_attn_out.shape[:-2], + core_attn_out.shape[-2] * core_attn_out.shape[-1], + ) + + output, _ = self.out_proj(core_attn_out) + return output + + +class Qwen3_5LinearDecoderLayer(nn.Module): + """Qwen3.5 Decoder Layer with Linear Attention (GatedDeltaNet).""" + + def __init__( + self, + config: Qwen3_5TextConfig, + layer_id: int, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + alt_stream: Optional[torch.cuda.Stream] = None, + is_nextn: bool = False, + ) -> None: + super().__init__() + self.config = config + self.layer_id = layer_id + + self.linear_attn = Qwen3_5GatedDeltaNet( + config, layer_id, quant_config, alt_stream, prefix + ) + + # NOTE: Determine the MLP type based on the model type + # Qwen3.5 use all layers for MLP / Qwen3.5-MoE use sparse MoE blocks + if config.model_type == "qwen3_5_moe_text": + self.mlp = Qwen2MoeSparseMoeBlock( + layer_id=layer_id, + config=config, + quant_config=quant_config, + alt_stream=( + alt_stream + if (_is_cuda or _disable_shared_experts_fusion()) + else None + ), + prefix=add_prefix("mlp", prefix.replace(".linear_attn", "")), + is_nextn=is_nextn, + support_shared_expert_fusion=not _disable_shared_experts_fusion(), + ) + is_layer_sparse = True + is_previous_layer_sparse = True + is_next_layer_sparse = True + elif config.model_type == "qwen3_5_text": + self.mlp = Qwen2MoeMLP( + hidden_size=config.hidden_size, + intermediate_size=config.intermediate_size, + hidden_act=config.hidden_act, + quant_config=quant_config, + prefix=add_prefix("mlp", prefix.replace(".linear_attn", "")), + ) + _maybe_enable_silu_fp4_quant_fusion(self.mlp) + is_layer_sparse = False + is_previous_layer_sparse = False + is_next_layer_sparse = False + else: + raise ValueError(f"Invalid model type: {config.model_type}") + + self.layer_scatter_modes = LayerScatterModes.init_new( + layer_id=layer_id, + num_layers=config.num_hidden_layers, + is_layer_sparse=is_layer_sparse, + is_previous_layer_sparse=is_previous_layer_sparse, + is_next_layer_sparse=is_next_layer_sparse, + ) + + self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = GemmaRMSNorm( + config.hidden_size, eps=config.rms_norm_eps + ) + # GDN layers need both bf16 (for the small in_proj_ba gating + # projection) and a quantized tuple only when in_proj_qkvz can consume + # it. Otherwise, stay on the plain AR+RMSNorm path. + enable_fused_ar_quant = ( + _enable_qwen35_fused_ar_quant() + and _linear_accepts_fp8_tuple(self.linear_attn.in_proj_qkvz) + ) + self.layer_communicator = _layer_communicator_class(config, is_nextn)( + layer_scatter_modes=self.layer_scatter_modes, + input_layernorm=self.input_layernorm, + post_attention_layernorm=self.post_attention_layernorm, + allow_reduce_scatter=True, + is_last_layer=(layer_id == config.num_hidden_layers - 1), + enable_fused_ar_quant=enable_fused_ar_quant, + fused_ar_quant_keep_bf16=enable_fused_ar_quant, + ) + + def forward( + self, + hidden_states: torch.Tensor, + residual: Optional[torch.Tensor], + **kwargs, + ): + forward_batch = kwargs.get("forward_batch", None) + + hidden_states, residual = ( + self.layer_communicator.prepare_attn_and_capture_last_layer_outputs( + hidden_states, + residual, + forward_batch, + captured_last_layer_outputs=kwargs.get( + "captured_last_layer_outputs", None + ), + ) + ) + + # fused AR+quant hands down a (fp8, scale) / (bf16, fp8, scale) tuple + hs = hidden_states[0] if isinstance(hidden_states, tuple) else hidden_states + if not forward_batch.forward_mode.is_idle() and hs.shape[0] > 0: + hidden_states = self.linear_attn( + hidden_states, + forward_batch, + ) + + # Fully Connected + hidden_states, residual = self.layer_communicator.prepare_mlp( + hidden_states, residual, forward_batch + ) + + mlp_reduce_scatter = self.layer_communicator.should_use_reduce_scatter( + forward_batch + ) + + fuse_mlp_allreduce = ( + self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer( + forward_batch + ) + ) + defer_moe_finalize = ( + fuse_mlp_allreduce + and isinstance(hidden_states, torch.Tensor) + and isinstance(self.mlp, Qwen2MoeSparseMoeBlock) + and hasattr(self.layer_communicator, "should_use_finalize") + and self.layer_communicator.should_use_finalize( + forward_batch, int(hidden_states.shape[0]) + ) + ) + if ( + fuse_mlp_allreduce + and self.layer_communicator.is_last_layer + and not defer_moe_finalize + ): + # The last layer has no prepare_attn consumer for deferred AllReduce; + # fall back before MLP so postprocess_layer performs the collective. + fuse_mlp_allreduce = False + with get_forward().scoped( + fuse_mlp_allreduce=fuse_mlp_allreduce, + mlp_reduce_scatter=mlp_reduce_scatter, + ): + if isinstance(self.mlp, Qwen2MoeSparseMoeBlock): + hidden_states = self.mlp( + hidden_states, + forward_batch, + defer_finalize=defer_moe_finalize, + ) + else: + hidden_states = self.mlp(hidden_states) + if fuse_mlp_allreduce: + hidden_states = _finish_mlp_output( + hidden_states, expect_deferred=defer_moe_finalize + ) + else: + hidden_states, residual = self.layer_communicator.postprocess_layer( + hidden_states, residual, forward_batch + ) + + return hidden_states, residual + + +class Qwen3_5AttentionDecoderLayer(nn.Module): + """Qwen3.5 Decoder Layer with Full Attention.""" + + def __init__( + self, + config: Qwen3_5TextConfig, + layer_id: int, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + alt_stream: Optional[torch.cuda.Stream] = None, + is_nextn: bool = False, + ) -> None: + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.attn_tp_rank = get_parallel().attn_tp_rank + self.attn_tp_size = get_parallel().attn_tp_size + # A Qwen3.5 draft is rewritten to the MTP arch (model_config._config_draft_model), + # so is_nextn marks it. Drafts are TP-sharded and do not replicate KV under DCP. + dcp_size = 1 if is_nextn else get_parallel().attn_dcp_size + self.kv_tp_size = self.attn_tp_size // dcp_size + self.kv_tp_rank = self.attn_tp_rank // dcp_size + self.total_num_heads = config.num_attention_heads + assert self.total_num_heads % self.attn_tp_size == 0 + self.num_heads = self.total_num_heads // self.attn_tp_size + self.total_num_kv_heads = config.num_key_value_heads + if self.total_num_kv_heads >= self.kv_tp_size: + assert self.total_num_kv_heads % self.kv_tp_size == 0 + else: + assert self.kv_tp_size % self.total_num_kv_heads == 0 + self.num_kv_heads = max(1, self.total_num_kv_heads // self.kv_tp_size) + self.head_dim = config.head_dim or (self.hidden_size // self.num_heads) + self.q_size = self.num_heads * self.head_dim + self.kv_size = self.num_kv_heads * self.head_dim + self.scaling = self.head_dim**-0.5 + self.max_position_embeddings = getattr(config, "max_position_embeddings", 8192) + + self.rope_theta, rope_scaling = get_rope_config(config) + self.partial_rotary_factor = getattr(config, "partial_rotary_factor", 1.0) + self.layer_id = layer_id + + # If rope_scaling doesn't specify a scaling type, treat as no scaling + if rope_scaling and not ("rope_type" in rope_scaling or "type" in rope_scaling): + rope_scaling = None + + self.attn_output_gate = getattr(config, "attn_output_gate", True) + if self.attn_output_gate: + logger.warning_once("using attn output gate!") + + self.rotary_emb = get_rope( + head_size=self.head_dim, + rotary_dim=self.head_dim, + max_position=self.max_position_embeddings, + rope_scaling=rope_scaling, + base=self.rope_theta, + partial_rotary_factor=self.partial_rotary_factor, + is_neox_style=True, + dtype=torch.get_default_dtype(), + ) + + # Q stays sharded across attention TP ranks; K/V are replicated within + # each DCP group. + self.qkv_proj = QKVParallelLinear( + config.hidden_size, + self.head_dim, + self.total_num_heads * (1 + self.attn_output_gate), + self.total_num_kv_heads, + bias=False, + quant_config=quant_config, + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + kv_tp_rank=self.kv_tp_rank, + kv_tp_size=self.kv_tp_size, + prefix=add_prefix("qkv_proj", prefix), + ) + + self.o_proj = RowParallelLinear( + self.total_num_heads * self.head_dim, + config.hidden_size, + bias=False, + quant_config=quant_config, + reduce_results=False, + tp_rank=self.attn_tp_rank, + tp_size=self.attn_tp_size, + prefix=add_prefix("o_proj", prefix), + ) + + self.attn = RadixAttention( + self.num_heads, + self.head_dim, + self.scaling, + num_kv_heads=self.num_kv_heads, + layer_id=layer_id, + prefix=f"{prefix}.attn", + quant_config=quant_config, + ) + + # Dense MLP for non-MoE variant + if config.model_type == "qwen3_5_text": + self.mlp = Qwen2MoeMLP( + hidden_size=config.hidden_size, + intermediate_size=config.intermediate_size, + hidden_act=config.hidden_act, + quant_config=quant_config, + prefix=add_prefix("mlp", prefix.replace(".self_attn", "")), + ) + is_layer_sparse = False + is_previous_layer_sparse = False + is_next_layer_sparse = False + elif config.model_type == "qwen3_5_moe_text": + self.mlp = Qwen2MoeSparseMoeBlock( + layer_id=layer_id, + config=config, + quant_config=quant_config, + alt_stream=( + alt_stream + if (_is_cuda or _disable_shared_experts_fusion()) + else None + ), + prefix=add_prefix("mlp", prefix.replace(".self_attn", "")), + is_nextn=is_nextn, + support_shared_expert_fusion=not _disable_shared_experts_fusion(), + ) + is_layer_sparse = True + is_previous_layer_sparse = True + is_next_layer_sparse = True + else: + raise ValueError(f"Invalid model type: {config.model_type}") + + self.layer_scatter_modes = LayerScatterModes.init_new( + layer_id=layer_id, + num_layers=config.num_hidden_layers, + is_layer_sparse=is_layer_sparse, + is_previous_layer_sparse=is_previous_layer_sparse, + is_next_layer_sparse=is_next_layer_sparse, + ) + + self.input_layernorm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = GemmaRMSNorm( + config.hidden_size, eps=config.rms_norm_eps + ) + + self.q_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps) + self.k_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps) + + # Standard attention layers benefit from a fused quant epilogue only + # when qkv_proj can consume the returned quantized tuple. + enable_fused_ar_quant = ( + _enable_qwen35_fused_ar_quant() and _linear_accepts_fp8_tuple(self.qkv_proj) + ) + self.layer_communicator = _layer_communicator_class(config, is_nextn)( + layer_scatter_modes=self.layer_scatter_modes, + input_layernorm=self.input_layernorm, + post_attention_layernorm=self.post_attention_layernorm, + allow_reduce_scatter=True, + is_last_layer=(layer_id == config.num_hidden_layers - 1), + enable_fused_ar_quant=enable_fused_ar_quant, + fused_ar_quant_keep_bf16=False, + ) + + self.alt_stream = alt_stream + + def _apply_qk_norm( + self, q: torch.Tensor, k: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply Q/K normalization with optional alt_stream overlap.""" + if ( + self.alt_stream is not None + and get_is_capture_mode() + and _qknorm_use_alt_stream + ): + current_stream = torch.cuda.current_stream() + self.alt_stream.wait_stream(current_stream) + q_by_head = q.reshape(-1, self.head_dim) + q_by_head = self.q_norm(q_by_head) + with torch.cuda.stream(self.alt_stream): + k_by_head = k.reshape(-1, self.head_dim) + k_by_head = self.k_norm(k_by_head) + current_stream.wait_stream(self.alt_stream) + elif _is_hip or _is_xpu or _is_cpu: + q_by_head, k_by_head = fused_qk_gemma_rmsnorm( + q, + k, + self.q_norm.weight.data, + self.k_norm.weight.data, + self.q_norm.variance_epsilon, + self.head_dim, + ) + else: + q_by_head = q.reshape(-1, self.head_dim) + q_by_head = self.q_norm(q_by_head) + k_by_head = k.reshape(-1, self.head_dim) + k_by_head = self.k_norm(k_by_head) + q = q_by_head.view(q.shape) + k = k_by_head.view(k.shape) + return q, k + + def forward_prepare_cuda_fused(self, positions, hidden_states): + """Fused QK GemmaRMSNorm + NeoX RoPE + gate deinterleave.""" + qkv, _ = self.qkv_proj(hidden_states) + if self.attn_output_gate: + q_gate, k, v = qkv.split( + [self.q_size * 2, self.kv_size, self.kv_size], dim=-1 + ) + else: + q_gate, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + q_out, k_out, gate_out = fused_qk_gemma_rmsnorm_rope_gate( + q_gate, + k, + self.q_norm.weight.data, + self.k_norm.weight.data, + self.rotary_emb.cos_sin_cache, + positions, + self.q_norm.variance_epsilon, + self.num_heads, + self.num_kv_heads, + self.head_dim, + self.rotary_emb.rotary_dim, + has_gate=self.attn_output_gate, + mrope_axis_map=(self.rotary_emb.axis_map if positions.dim() == 2 else None), + ) + seq_len = hidden_states.shape[0] + q = q_out.view(seq_len, -1) + k = k_out.view(seq_len, -1) + gate = gate_out.view(seq_len, -1) if gate_out is not None else None + return q, k, v, gate + + def forward_prepare_native(self, positions, hidden_states): + if _use_aiter and isinstance(hidden_states, tuple): + hidden_states = _select_fused_ar_input_for_linear( + hidden_states, self.qkv_proj + ) + qkv, _ = self.qkv_proj(hidden_states) + if self.attn_output_gate: + q_gate, k, v = qkv.split( + [self.q_size * 2, self.kv_size, self.kv_size], dim=-1 + ) + orig_shape = q_gate.shape[:-1] + q_gate = q_gate.view(*orig_shape, self.num_heads, -1) + q, gate = torch.chunk(q_gate, 2, dim=-1) + q = q.reshape(*orig_shape, -1) + # gate stays as 3D strided view; fused_sigmoid_mul handles it directly + else: + q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + gate = None + + q, k = self._apply_qk_norm(q, k) + q, k = self.rotary_emb(positions, q, k) + return q, k, v, gate + + def forward_prepare_fused_gate(self, positions, hidden_states): + if _use_aiter and isinstance(hidden_states, tuple): + hidden_states = _select_fused_ar_input_for_linear( + hidden_states, self.qkv_proj + ) + qkv, _ = self.qkv_proj(hidden_states) + if self.attn_output_gate: + q_gate, k, v = qkv.split( + [self.q_size * 2, self.kv_size, self.kv_size], dim=-1 + ) + seq_len = q_gate.shape[0] + q_flat, k_flat, gate_flat = fused_qk_gemma_rmsnorm_with_gate( + q_gate, + k, + self.q_norm.weight.data, + self.k_norm.weight.data, + self.q_norm.variance_epsilon, + self.head_dim, + self.num_heads, + ) + q = q_flat.view(seq_len, -1) + k = k_flat.view(seq_len, -1) + gate = gate_flat.view(seq_len, -1) + else: + q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + gate = None + q, k = self._apply_qk_norm(q, k) + + q, k = self.rotary_emb(positions, q, k) + return q, k, v, gate + + def forward_prepare_npu(self, positions, hidden_states, forward_batch): + qkv, _ = self.qkv_proj(hidden_states) + # Calculate first full attention layer ID based on config + if self.attn.layer_id == (self.config.full_attention_interval - 1): + self.rotary_emb.get_cos_sin_with_position(positions) + + q, k, v, gate = split_qkvgate_gemma_rmsnorm_rope( + qkv, + self.rotary_emb.position_sin, + self.rotary_emb.position_cos, + self.q_size, + self.kv_size, + self.head_dim, + int(self.head_dim * self.partial_rotary_factor), + eps=self.q_norm.variance_epsilon, + q_weight=self.q_norm.weight, + k_weight=self.k_norm.weight, + ) + return q, k, v, gate + + def self_attention( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + forward_batch: ForwardBatch, + ) -> torch.Tensor: + """Full attention forward pass.""" + if _is_cuda and self.attn_output_gate: + q, k, v, gate = self.forward_prepare_cuda_fused( + positions=positions, + hidden_states=hidden_states, + ) + elif (_is_hip or _is_xpu or _is_cpu) and self.attn_output_gate: + q, k, v, gate = self.forward_prepare_fused_gate( + positions=positions, + hidden_states=hidden_states, + ) + elif ( + not _is_npu + or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed() + or not self.attn_output_gate + ): + q, k, v, gate = self.forward_prepare_native( + positions=positions, + hidden_states=hidden_states, + ) + else: + q, k, v, gate = self.forward_prepare_npu( + positions=positions, + hidden_states=hidden_states, + forward_batch=forward_batch, + ) + + attn_output = self.attn(q, k, v, forward_batch) + + if self.attn_output_gate: + if not _is_npu: + attn_output = fused_sigmoid_mul(attn_output, gate, inplace=True) + else: + gate_val = gate.reshape(gate.shape[0], -1) if gate.ndim == 3 else gate + attn_output.mul_(torch.sigmoid(gate_val)) + + output, _ = self.o_proj(attn_output) + return output + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + residual: Optional[torch.Tensor], + forward_batch: ForwardBatch, + captured_last_layer_outputs: Optional[list[torch.Tensor]] = None, + **kwargs, + ): + hidden_states, residual = ( + self.layer_communicator.prepare_attn_and_capture_last_layer_outputs( + hidden_states, + residual, + forward_batch, + captured_last_layer_outputs=captured_last_layer_outputs, + ) + ) + + # fused AR+quant hands down a (fp8, scale) / (bf16, fp8, scale) tuple + hs = hidden_states[0] if isinstance(hidden_states, tuple) else hidden_states + if not forward_batch.forward_mode.is_idle() and hs.shape[0] > 0: + hidden_states = self.self_attention( + positions=positions, + hidden_states=hidden_states, + forward_batch=forward_batch, + ) + + # Fully Connected + hidden_states, residual = self.layer_communicator.prepare_mlp( + hidden_states, residual, forward_batch + ) + mlp_reduce_scatter = self.layer_communicator.should_use_reduce_scatter( + forward_batch + ) + + fuse_mlp_allreduce = ( + self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer( + forward_batch + ) + ) + defer_moe_finalize = ( + fuse_mlp_allreduce + and isinstance(hidden_states, torch.Tensor) + and isinstance(self.mlp, Qwen2MoeSparseMoeBlock) + and hasattr(self.layer_communicator, "should_use_finalize") + and self.layer_communicator.should_use_finalize( + forward_batch, int(hidden_states.shape[0]) + ) + ) + if ( + fuse_mlp_allreduce + and self.layer_communicator.is_last_layer + and not defer_moe_finalize + ): + fuse_mlp_allreduce = False + with get_forward().scoped( + fuse_mlp_allreduce=fuse_mlp_allreduce, + mlp_reduce_scatter=mlp_reduce_scatter, + ): + if isinstance(self.mlp, Qwen2MoeSparseMoeBlock): + hidden_states = self.mlp( + hidden_states, + forward_batch, + defer_finalize=defer_moe_finalize, + ) + else: + hidden_states = self.mlp(hidden_states) + if fuse_mlp_allreduce: + hidden_states = _finish_mlp_output( + hidden_states, expect_deferred=defer_moe_finalize + ) + else: + hidden_states, residual = self.layer_communicator.postprocess_layer( + hidden_states, residual, forward_batch + ) + + return hidden_states, residual + + +ALL_DECODER_LAYER_TYPES = { + "attention": Qwen3_5AttentionDecoderLayer, + "linear_attention": Qwen3_5LinearDecoderLayer, +} + +# ModelOpt FP4 checkpoints bake the per-layer KV-cache scales under the HF +# attention projections; in sglang they live on RadixAttention. Apply this to the +# weight stream at the top of load_weights(), before ".self_attn" is stripped and +# before the stacked qkv_proj matching would consume the name. + +# === agnes === +# Agnes 3.0 Flash checkpoints (config model_type "agnes") use their own tensor +# prefixes and carry a parallel FFN branch per layer. This generator sits at +# the top of the weight stream and turns it into what the implementation below +# expects: delta_attn -> linear_attn, global_attn -> self_attn, and the branch +# concatenated onto the main gate / up (dim 0) and down (dim 1) projections, +# matching the widened intermediate_size set by sglang.srt.configs.agnes. +import json as _agnes_json +import os as _agnes_os +import re as _agnes_re + +_AGNES_MLP_RE = _agnes_re.compile(r"^(.*\.layers\.\d+\.mlp\.)(gate_proj|up_proj|down_proj)\.weight$") + + +class _AgnesBranchReader: + def __init__(self, model_path): + from safetensors import safe_open + + self._open = safe_open + self.path = model_path + index = _agnes_os.path.join(model_path, "model.safetensors.index.json") + self.weight_map = _agnes_json.load(open(index))["weight_map"] + self.handles = {} + + def get(self, key): + fn = self.weight_map[key] + if fn not in self.handles: + self.handles[fn] = self._open(_agnes_os.path.join(self.path, fn), framework="pt", device="cpu") + return self.handles[fn].get_tensor(key) + + +def _agnes_translate_weights(model, weights): + cfg = getattr(model.config, "text_config", None) or model.config + width = int(getattr(cfg, "agnes_parallel_ffn_intermediate_size", 0) or 0) + if width <= 0: + yield from weights + return + model_path = ( + getattr(cfg, "agnes_model_path", None) + or getattr(model.config, "agnes_model_path", None) + or _agnes_os.environ.get("AGNES_MODEL_PATH") + or getattr(model.config, "_name_or_path", None) + ) + if not model_path or not _agnes_os.path.isdir(model_path): + raise RuntimeError( + f"agnes: cannot locate the checkpoint directory (got {model_path!r}); " + "set AGNES_MODEL_PATH to the model directory" + ) + reader = _AgnesBranchReader(model_path) + for name, w in weights: + if ".mlp.parallel_ffn." in name: + continue + m = _AGNES_MLP_RE.match(name) + if m and "visual" not in name and not name.startswith("mtp"): + extra = reader.get(f"{m.group(1)}parallel_ffn.{m.group(2)}.weight") + dim = 1 if m.group(2) == "down_proj" else 0 + w = torch.cat([w, extra.to(device=w.device, dtype=w.dtype)], dim=dim) + name = name.replace(".delta_attn.", ".linear_attn.").replace(".global_attn.", ".self_attn.") + yield name, w +# === /agnes === + +QWEN3_5_KV_SCALE_MAPPER = WeightsMapper( + orig_to_new_substr={ + ".self_attn.k_proj.k_scale": ".attn.k_scale", + ".self_attn.v_proj.v_scale": ".attn.v_scale", + # compressed-tensors stores kv_cache_scheme scales on the attention module. + ".self_attn.k_scale": ".attn.k_scale", + ".self_attn.v_scale": ".attn.v_scale", + }, +) + + +class Qwen3_5ForCausalLM(nn.Module): + """Qwen3.5 Model with support for dense variant.""" + + packed_modules_mapping = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], + "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"], + "in_proj_ba": ["in_proj_b", "in_proj_a"], + } + + supported_lora_modules = [ + "qkv_proj", + "o_proj", + "out_proj", + "in_proj_qkvz", + "gate_up_proj", + "down_proj", + "lm_head", + ] + + def get_hidden_dim(self, module_name: str, layer_idx: int): + config = self.config + head_dim = config.head_dim or (config.hidden_size // config.num_attention_heads) + + if module_name == "qkv_proj": + attn_output_gate = getattr(config, "attn_output_gate", True) + q_heads = config.num_attention_heads * (2 if attn_output_gate else 1) + return ( + config.hidden_size, + head_dim * (q_heads + config.num_key_value_heads * 2), + ) + elif module_name == "o_proj": + return config.num_attention_heads * head_dim, config.hidden_size + elif module_name == "out_proj": + value_dim = config.linear_value_head_dim * config.linear_num_value_heads + return value_dim, config.hidden_size + elif module_name == "in_proj_qkvz": + key_dim = config.linear_key_head_dim * config.linear_num_key_heads + value_dim = config.linear_value_head_dim * config.linear_num_value_heads + return config.hidden_size, key_dim * 2 + value_dim * 2 + elif module_name == "gate_up_proj": + # MoE: shared expert uses shared_expert_intermediate_size + # Dense: regular MLP uses intermediate_size + is_moe = "moe" in getattr(config, "model_type", "") + if is_moe: + inter = config.shared_expert_intermediate_size + else: + inter = config.intermediate_size + return config.hidden_size, inter * 2 + elif module_name == "down_proj": + is_moe = "moe" in getattr(config, "model_type", "") + if is_moe: + inter = config.shared_expert_intermediate_size + else: + inter = config.intermediate_size + return inter, config.hidden_size + elif module_name == "gate_up_proj_moe": + return config.hidden_size, config.moe_intermediate_size * 2 + elif module_name == "down_proj_moe": + return config.moe_intermediate_size, config.hidden_size + elif module_name == "embed_tokens": + return config.vocab_size, config.hidden_size + elif module_name == "lm_head": + return config.hidden_size, config.vocab_size + else: + raise NotImplementedError( + f"get_hidden_dim not implemented for {module_name}" + ) + + def __init__( + self, + config: Qwen3_5TextConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + is_nextn: bool = False, + ) -> None: + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.pp_group = get_pp_group() + + alt_stream = get_stream("alt") if _is_cuda or _hip_use_alt_stream else None + + # Embedding layer + if self.pp_group.is_first_rank: + self.embed_tokens = VocabParallelEmbedding( + config.vocab_size, + config.hidden_size, + org_num_embeddings=config.vocab_size, + enable_tp=not is_dp_attention_enabled(), + ) + else: + self.embed_tokens = PPMissingLayer() + + # Decoder layers + def get_layer(idx: int, prefix: str): + layer_type = config.layers_block_type[idx] + layer_class = ALL_DECODER_LAYER_TYPES[layer_type] + if layer_type == "attention": + prefix = add_prefix("self_attn", prefix) + else: + prefix = add_prefix("linear_attn", prefix) + return layer_class( + config=config, + layer_id=idx, + quant_config=quant_config, + prefix=prefix, + alt_stream=alt_stream, + is_nextn=is_nextn, + ) + + self.layers, self._start_layer, self._end_layer = make_layers( + config.num_hidden_layers, + get_layer, + pp_rank=self.pp_group.rank_in_group, + pp_size=self.pp_group.world_size, + prefix=f"{prefix}.layers", + ) + + self.flashinfer_mnnvl_cutedsl_fusion = None + if _use_mnnvl_cutedsl_fusion(config, is_nextn): + if self.pp_group.world_size != 1: + raise RuntimeError( + "Qwen3.5 FlashInfer MNNVL CuTe DSL fusion currently requires PP=1" + ) + unsupported_layers = [ + layer.layer_id + for layer in self.layers + if not isinstance(layer.mlp, Qwen2MoeSparseMoeBlock) + or not layer.mlp.supports_deferred_finalize + ] + if unsupported_layers: + raise RuntimeError( + "Qwen3.5 FlashInfer MNNVL CuTe DSL fusion currently " + "requires block-FP8 MoE weights with FlashInfer TRTLLM " + "deferred-finalize support on every layer; unsupported " + "layers: " + f"{unsupported_layers}" + ) + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + Qwen35FlashInferFusionService, + Qwen35FlashInferLayerCommunicator, + ) + + self.flashinfer_mnnvl_cutedsl_fusion = Qwen35FlashInferFusionService( + hidden_size=config.hidden_size, + top_k=config.num_experts_per_tok, + rms_epsilon=config.rms_norm_eps, + ) + for layer in self.layers: + communicator = layer.layer_communicator + if not isinstance(communicator, Qwen35FlashInferLayerCommunicator): + raise RuntimeError( + "Qwen3.5 fusion-enabled layer has the wrong communicator" + ) + communicator.fusion_service = self.flashinfer_mnnvl_cutedsl_fusion + logger.info( + "Installed one Qwen3.5 FlashInfer fusion handle for %d layers", + len(self.layers), + ) + + # Final normalization + if self.pp_group.is_last_rank: + self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + else: + self.norm = PPMissingLayer() + + self.layers_to_capture = [] + + def get_input_embeddings(self): + return self.embed_tokens + + def prepare_before_cuda_graph_capture(self, model_runner) -> None: + if self.flashinfer_mnnvl_cutedsl_fusion is None: + return + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + prepare_qwen35_flashinfer_fusion, + ) + + prepare_qwen35_flashinfer_fusion(self, model_runner) + + def set_dflash_layers_to_capture(self, layers_to_capture: list[int]): + self.layers_to_capture = layers_to_capture + for layer_id in self.layers_to_capture: + setattr(self.layers[layer_id], "_is_layer_to_capture", True) + + @property + def start_layer(self) -> int: + return self._start_layer + + @property + def end_layer(self) -> int: + return self._end_layer + + @torch.no_grad() + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + forward_batch: ForwardBatch, + input_embeds: Optional[torch.Tensor] = None, + pp_proxy_tensors: Optional[PPProxyTensors] = None, + input_deepstack_embeds: Optional[torch.Tensor] = None, + ) -> Union[torch.Tensor, PPProxyTensors]: + if ( + self.flashinfer_mnnvl_cutedsl_fusion is not None + and input_deepstack_embeds is not None + and input_deepstack_embeds.numel() > 0 + ): + raise RuntimeError( + "Qwen3.5 FlashInfer MNNVL CuTe DSL fusion currently supports " + "the text-only path, not deepstack visual inputs" + ) + + # Initialize hidden states + if self.pp_group.is_first_rank: + if input_embeds is None: + hidden_states = self.embed_tokens(input_ids) + else: + hidden_states = input_embeds + residual = None + else: + assert pp_proxy_tensors is not None + hidden_states = pp_proxy_tensors["hidden_states"] + residual = pp_proxy_tensors["residual"] + + aux_hidden_states = [] + # Pass through decoder layers + for layer_idx in range(self.start_layer, self.end_layer): + layer = self.layers[layer_idx] + with get_global_expert_distribution_recorder().with_current_layer( + layer_idx + ): + hidden_states, residual = layer( + positions=positions, + hidden_states=hidden_states, + residual=residual, + forward_batch=forward_batch, + captured_last_layer_outputs=( + aux_hidden_states + if getattr(layer, "_is_layer_to_capture", False) + else None + ), + ) + + # Process deepstack embeddings if provided + if ( + input_deepstack_embeds is not None + and input_deepstack_embeds.numel() > 0 + and layer_idx < 3 + ): + sep = self.hidden_size * layer_idx + hidden_states.add_( + input_deepstack_embeds[:, sep : sep + self.hidden_size] + ) + + # Return intermediate tensors for pipeline parallelism + if not self.pp_group.is_last_rank: + return PPProxyTensors( + { + "hidden_states": hidden_states, + "residual": residual, + } + ) + + # The final layer has no successor to consume its deferred MoE tail. + trace_final_norm = envs.SGLANG_TRACE_QWEN35_FINAL_NORM.get() + use_native_final_norm = envs.SGLANG_QWEN35_NATIVE_FINAL_NORM.get() + is_deferred_finalize = False + if self.flashinfer_mnnvl_cutedsl_fusion is not None: + from sglang.srt.layers.moe.qwen35_flashinfer_fusion import ( + Qwen35MoeFinalizeHandoff, + ) + + is_deferred_finalize = isinstance(hidden_states, Qwen35MoeFinalizeHandoff) + + if is_deferred_finalize: + if residual is None or self.flashinfer_mnnvl_cutedsl_fusion is None: + raise RuntimeError("invalid final deferred MoE handoff") + hidden_states, _ = self.flashinfer_mnnvl_cutedsl_fusion.finalize( + hidden_states, residual, self.norm.gemma_weight + ) + elif hidden_states.shape[0] != 0: + if trace_final_norm: + print( + "SGLANG_TRACE_QWEN35_FINAL_NORM " + f"stage=pre_sync_enter hidden={tuple(hidden_states.shape)} " + f"hidden_stride={hidden_states.stride()} " + f"hidden_dtype={hidden_states.dtype} " + f"hidden_contiguous={hidden_states.is_contiguous()} " + f"residual={None if residual is None else tuple(residual.shape)} " + f"native={use_native_final_norm}", + flush=True, + ) + torch.cuda.synchronize() + print( + "SGLANG_TRACE_QWEN35_FINAL_NORM stage=pre_sync_returned", + flush=True, + ) + if residual is None: + hidden_states = ( + self.norm.forward_native(hidden_states) + if use_native_final_norm + else self.norm(hidden_states) + ) + else: + hidden_states, _ = ( + self.norm.forward_native(hidden_states, residual) + if use_native_final_norm + else self.norm(hidden_states, residual) + ) + if trace_final_norm: + print( + "SGLANG_TRACE_QWEN35_FINAL_NORM stage=post_sync_enter", + flush=True, + ) + torch.cuda.synchronize() + print( + "SGLANG_TRACE_QWEN35_FINAL_NORM stage=post_sync_returned", + flush=True, + ) + + if len(aux_hidden_states) == 0: + return hidden_states + + return hidden_states, aux_hidden_states + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + weights = _agnes_translate_weights(self, weights) + weights = QWEN3_5_KV_SCALE_MAPPER.apply(weights) + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + # GDN + ("in_proj_qkvz.", "in_proj_qkv.", (0, 1, 2)), + ("in_proj_qkvz.", "in_proj_z.", 3), + ("in_proj_ba.", "in_proj_b.", 0), + ("in_proj_ba.", "in_proj_a.", 1), + ] + + loaded_params: Set[str] = set() + params_dict = dict(self.named_parameters(remove_duplicate=False)) + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + if "mtp" in name: + continue + if "visual" in name: + continue + if "language_model" in name: + name = name.replace(r"model.language_model.", r"model.") + if ".self_attn." in name: + name = name.replace(".self_attn", "") + layer_id = get_layer_id(name) + if ( + layer_id is not None + and hasattr(self, "start_layer") + and (layer_id < self.start_layer or layer_id >= self.end_layer) + ): + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + + if "mlp.experts" in name: + continue + + name = name.replace(weight_name, param_name) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + # Skip layers on other devices. + # if is_pp_missing_parameter(name, self): + # continue + if name not in params_dict: + continue + param = params_dict[name] + weight_loader = getattr(param, "weight_loader") + weight_loader(param, loaded_weight, shard_id) + break + else: + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + if name not in params_dict: + logger.warning(f"Parameter {name} not found in params_dict") + continue + param = params_dict[name] + + weight_loader = getattr(param, "weight_loader", default_weight_loader) + weight_loader(param, loaded_weight) + loaded_params.add(name) + return loaded_params + + @classmethod + def get_model_config_for_expert_location(cls, config): + return ModelConfigForExpertLocation( + num_layers=config.num_hidden_layers, + num_logical_experts=config.num_experts, + num_groups=None, + ) + + +class Qwen3_5MoeForCausalLM(Qwen3_5ForCausalLM): + def __init__( + self, + config: Qwen3_5TextConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + ) -> None: + super().__init__(config=config, quant_config=quant_config, prefix=prefix) + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + weights = _agnes_translate_weights(self, weights) + weights = QWEN3_5_KV_SCALE_MAPPER.apply(weights) + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + # GDN + ("in_proj_qkvz.", "in_proj_qkv.", (0, 1, 2)), + ("in_proj_qkvz.", "in_proj_z.", 3), + ("in_proj_ba.", "in_proj_b.", 0), + ("in_proj_ba.", "in_proj_a.", 1), + ] + + # Params for weights, fp8 weight scales, fp8 activation scales + # (param_name, weight_name, expert_id, shard_id) + expert_params_mapping = FusedMoE.make_expert_params_mapping( + ckpt_gate_proj_name="gate_proj", + ckpt_down_proj_name="down_proj", + ckpt_up_proj_name="up_proj", + num_experts=self.config.num_experts, + ) + + # Skip loading extra parameters for GPTQ/modelopt models. + ignore_suffixes = ( + ".bias", + "_bias", + ".k_scale", + "_k_scale", + ".v_scale", + "_v_scale", + ".weight_scale", + "_weight_scale", + ".input_scale", + "_input_scale", + ) + + is_fused_expert = False + fused_expert_params_mapping = [ + ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"), + ("experts.w2_weight", "experts.down_proj", 0, "w2"), + ] + + num_experts = self.config.num_experts + + def load_fused_expert_weights( + name: str, + params_dict: dict, + loaded_weight: torch.Tensor, + shard_id: str, + num_experts: int, + ): + if name not in params_dict: + return False + param = params_dict[name] + weight_loader = param.weight_loader + # let ep moe layer to gracefully handle expert_ids that do not belong to local moe rank + for expert_id in range(num_experts): + curr_expert_weight = loaded_weight[expert_id] + weight_loader( + param, + curr_expert_weight, + name, + shard_id, + expert_id, + ) + return True + + loaded_params: Set[str] = set() + params_dict = dict(self.named_parameters(remove_duplicate=False)) + + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + if "mtp" in name: + continue + if "visual" in name: + continue + if "language_model" in name: + name = name.replace(r"model.language_model.", r"model.") + if ".self_attn." in name: + name = name.replace(".self_attn", "") + + layer_id = get_layer_id(name) + if ( + layer_id is not None + and hasattr(self, "start_layer") + and (layer_id < self.start_layer or layer_id >= self.end_layer) + ): + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if "experts.gate_up_proj" in name or "experts.down_proj" in name: + is_fused_expert = True + expert_params_mapping = fused_expert_params_mapping + + # Skip non-stacked layers and experts (experts handled below). + if weight_name not in name: + continue + + # We have mlp.experts[0].gate_proj in the checkpoint. + # Since we handle the experts below in expert_params_mapping, + # we need to skip here BEFORE we update the name, otherwise + # name will be updated to mlp.experts[0].gate_up_proj, which + # will then be updated below in expert_params_mapping + # for mlp.experts[0].gate_gate_up_proj, which breaks load. + if "mlp.experts" in name: + continue + name = name.replace(weight_name, param_name) + # Skip loading extra parameters for GPTQ/modelopt models. + if name.endswith(ignore_suffixes) and name not in params_dict: + continue + + if name not in params_dict: + continue + + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + # Track if this is an expert weight to enable early skipping + is_expert_weight = False + + for mapping in expert_params_mapping: + param_name, weight_name, expert_id, shard_id = mapping + if weight_name not in name: + continue + # Anyway, this is an expert weight and should not be + # attempted to load as other weights later + is_expert_weight = True + name_mapped = name.replace(weight_name, param_name) + if is_fused_expert: + if "experts.gate_up_proj" in name: + loaded_weight = loaded_weight.chunk(2, dim=-2) + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight[0], + "w1", + num_experts, + ) + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight[1], + "w3", + num_experts, + ) + else: + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight, + shard_id, + num_experts, + ) + else: + # Skip loading extra parameters for GPTQ/modelopt models. + if ( + name_mapped.endswith(ignore_suffixes) + and name_mapped not in params_dict + ): + continue + param = params_dict[name_mapped] + # We should ask the weight loader to return success or + # not here since otherwise we may skip experts with + # # other available replicas. + weight_loader = param.weight_loader + weight_loader( + param, + loaded_weight, + name_mapped, + shard_id=shard_id, + expert_id=expert_id, + ) + name = name_mapped + break + else: + if is_expert_weight: + # This is an expert weight but not mapped to this rank, skip all remaining processing + continue + + # Skip loading extra parameters for GPTQ/modelopt models. + if name.endswith(ignore_suffixes) and name not in params_dict: + continue + + if name in params_dict.keys(): + param = params_dict[name] + weight_loader = getattr( + param, "weight_loader", default_weight_loader + ) + weight_loader(param, loaded_weight) + else: + logger.warning(f"Parameter {name} not found in params_dict") + loaded_params.add(name) + + return loaded_params + + +class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration): + packed_modules_mapping = Qwen3_5ForCausalLM.packed_modules_mapping + hf_to_sglang_mapper = None + + supported_lora_modules = Qwen3_5ForCausalLM.supported_lora_modules + + def __init__( + self, + config: Qwen3_5Config, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + language_model_cls=Qwen3_5ForCausalLM, + ): + super().__init__(config, quant_config, prefix, language_model_cls) + + rope_config = getattr(self.config, "rope_parameters", None) or getattr( + self.config, "rope_scaling", {} + ) + self.is_mrope_enabled = ( + not self.language_model_only and "mrope_section" in rope_config + ) + + self.deepstack_visual_indexes = ( + self.visual.deepstack_visual_indexes if self.visual is not None else [] + ) + + def get_hidden_dim(self, module_name: str, layer_idx: int): + return self.model.get_hidden_dim(module_name, layer_idx) + + def should_apply_lora(self, module_name: str) -> bool: + return module_name.startswith("model.layers.") + + @property + def start_layer(self) -> int: + return getattr(getattr(self, "model", None), "start_layer", 0) + + @property + def end_layer(self) -> int: + model = getattr(self, "model", None) + end_layer = getattr(model, "end_layer", None) + if end_layer is not None: + return end_layer + cfg = getattr(model, "config", None) + return int(getattr(cfg, "num_hidden_layers", 0)) + + def get_embed_and_head(self): + embed = self.model.embed_tokens.weight if self.pp_group.is_first_rank else None + head = self.lm_head.weight if self.pp_group.is_last_rank else None + return embed, head + + def set_embed_and_head(self, embed, head): + if self.pp_group.is_first_rank and embed is not None: + del self.model.embed_tokens.weight + self.model.embed_tokens.weight = embed + if self.pp_group.is_last_rank and head is not None: + del self.lm_head.weight + self.lm_head.weight = head + if _is_xpu: + torch.xpu.empty_cache() + torch.xpu.synchronize() + else: + torch.cuda.empty_cache() + torch.cuda.synchronize() + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + weights = _agnes_translate_weights(self, weights) + weights = QWEN3_5_KV_SCALE_MAPPER.apply(weights) + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + # GDN fused projections + ("in_proj_qkvz.", "in_proj_qkv.", (0, 1, 2)), + ("in_proj_qkvz.", "in_proj_z.", 3), + ("in_proj_ba.", "in_proj_b.", 0), + ("in_proj_ba.", "in_proj_a.", 1), + ] + + loaded_params: Set[str] = set() + params_dict = dict(self.named_parameters(remove_duplicate=False)) + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + if "mtp" in name: + continue + if "language_model" in name: + name = name.replace(r"model.language_model.", r"model.") + if ".self_attn." in name: + name = name.replace(".self_attn", "") + if ( + self.config.tie_word_embeddings + and self.pp_group.is_last_rank + and "model.embed_tokens.weight" in name + ): + if "lm_head.weight" in params_dict: + lm_head_param = params_dict["lm_head.weight"] + weight_loader = getattr( + lm_head_param, "weight_loader", default_weight_loader + ) + weight_loader(lm_head_param, loaded_weight) + layer_id = get_layer_id(name) + if ( + layer_id is not None + and hasattr(self, "start_layer") + and (layer_id < self.start_layer or layer_id >= self.end_layer) + ): + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + + if "visual" in name or "mlp.experts" in name: + continue + + name = name.replace(weight_name, param_name) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + # Skip layers on other devices. + # if is_pp_missing_parameter(name, self): + # continue + if name not in params_dict: + continue + param = params_dict[name] + weight_loader = getattr(param, "weight_loader") + weight_loader(param, loaded_weight, shard_id) + break + else: + if "visual" in name: + # adapt to VisionAttention + name = name.replace(r"attn.qkv.", r"attn.qkv_proj.") + name = name.replace(r"model.visual.", r"visual.") + + # print(name, loaded_weight.shape) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + if name not in params_dict: + logger.warning(f"Parameter {name} not found in params_dict") + continue + param = params_dict[name] + + weight_loader = getattr(param, "weight_loader", default_weight_loader) + weight_loader(param, loaded_weight) + if ( + self.config.tie_word_embeddings + and name == "model.embed_tokens.weight" + and (_is_cpu and _is_amx_available) + ): + param_lm_head = params_dict["lm_head.weight"] + weight_loader = getattr( + param_lm_head, "weight_loader", default_weight_loader + ) + weight_loader(param_lm_head, loaded_weight) + loaded_params.add(name) + return loaded_params + + +class Qwen3_5MoeForConditionalGeneration(Qwen3VLForConditionalGeneration): + """Qwen3.5 MoE Vision-Language Model.""" + + packed_modules_mapping = Qwen3_5ForCausalLM.packed_modules_mapping + hf_to_sglang_mapper = None + + supported_lora_modules = Qwen3_5ForCausalLM.supported_lora_modules + + def __init__( + self, + config: Qwen3_5MoeConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + language_model_cls=Qwen3_5MoeForCausalLM, + ) -> None: + super().__init__(config, quant_config, prefix, language_model_cls) + rope_config = getattr(self.config, "rope_parameters", None) or getattr( + self.config, "rope_scaling", {} + ) + self.is_mrope_enabled = ( + not self.language_model_only and "mrope_section" in rope_config + ) + + self.deepstack_visual_indexes = ( + self.visual.deepstack_visual_indexes if self.visual is not None else [] + ) + + self.num_fused_shared_experts = 0 + if _use_aiter and not _disable_shared_experts_fusion(): + self.num_fused_shared_experts = self._get_num_fused_shared_experts() + + self.enable_shared_expert_fusion = self.num_fused_shared_experts > 0 + + def get_hidden_dim(self, module_name: str, layer_idx: int): + return self.model.get_hidden_dim(module_name, layer_idx) + + def prepare_before_cuda_graph_capture(self, model_runner) -> None: + prepare = getattr(self.model, "prepare_before_cuda_graph_capture", None) + if prepare is not None: + prepare(model_runner) + + def should_apply_lora(self, module_name: str) -> bool: + # Accept all language model layer modules (attention, linear_attn, mlp). + return module_name.startswith("model.layers.") + + def _get_num_fused_shared_experts(self): + if not hasattr(self.model, "layers"): + return 0 + for layer_id in range(self.model.start_layer, self.model.end_layer): + mlp = getattr(self.model.layers[layer_id], "mlp", None) + if hasattr(mlp, "num_fused_shared_experts"): + return mlp.num_fused_shared_experts + return 0 + + def get_embed_and_head(self): + embed = self.model.embed_tokens.weight if self.pp_group.is_first_rank else None + head = self.lm_head.weight if self.pp_group.is_last_rank else None + return embed, head + + def set_embed_and_head(self, embed, head): + if self.pp_group.is_first_rank and embed is not None: + del self.model.embed_tokens.weight + self.model.embed_tokens.weight = embed + if self.pp_group.is_last_rank and head is not None: + del self.lm_head.weight + self.lm_head.weight = head + if _is_xpu: + torch.xpu.empty_cache() + torch.xpu.synchronize() + else: + torch.cuda.empty_cache() + torch.cuda.synchronize() + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + weights = _agnes_translate_weights(self, weights) + weights = QWEN3_5_KV_SCALE_MAPPER.apply(weights) + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + # GDN fused projections + ("in_proj_qkvz.", "in_proj_qkv.", (0, 1, 2)), + ("in_proj_qkvz.", "in_proj_z.", 3), + ("in_proj_ba.", "in_proj_b.", 0), + ("in_proj_ba.", "in_proj_a.", 1), + ] + + num_experts = self.config.num_experts + + # Params for weights, fp8 weight scales, fp8 activation scales + # (param_name, weight_name, expert_id, shard_id) + expert_params_mapping = FusedMoE.make_expert_params_mapping( + ckpt_gate_proj_name="gate_proj", + ckpt_down_proj_name="down_proj", + ckpt_up_proj_name="up_proj", + num_experts=( + num_experts + if not self.enable_shared_expert_fusion + else num_experts + self.num_fused_shared_experts + ), + ) + + # Skip loading extra parameters for GPTQ/modelopt models. + ignore_suffixes = ( + ".bias", + "_bias", + ".k_scale", + "_k_scale", + ".v_scale", + "_v_scale", + "_weight_scale", + "_input_scale", + ) + + is_fused_expert = False + fused_expert_params_mapping = [ + ("experts.w13_weight", "experts.gate_up_proj", 0, "w1"), + ("experts.w2_weight", "experts.down_proj", 0, "w2"), + ] + + if self.enable_shared_expert_fusion: + """ + When shared experts are fused, we need to map the shared experts to routed experts. + + mlp.share_expert.gate_up_proj.weight --> experts.512.gate_up_proj.weight -> experts.w13_weight, expert_id = 512 + mlp.share_expert.down_proj.weight --> experts.512.down_proj.weight -> experts.w2_weight, expert_id = 512 + """ + fused_expert_params_mapping += [ + ( + "experts.w13_", + f"experts.{num_experts}.gate_up_proj.", + num_experts, + "w1", + ), + ( + "experts.w2_", + f"experts.{num_experts}.down_proj.", + num_experts, + "w2", + ), + ## shared experts may contain gate_proj and up_proj instead of gate_up_proj + ( + "experts.w13_", + f"experts.{num_experts}.gate_proj.", + num_experts, + "w1", + ), + ( + "experts.w13_", + f"experts.{num_experts}.up_proj.", + num_experts, + "w3", + ), + ] + + def load_fused_expert_weights( + name: str, + params_dict: dict, + loaded_weight: torch.Tensor, + shard_id: str, + num_experts: int, + ): + if name not in params_dict: + return False + param = params_dict[name] + weight_loader = param.weight_loader + # let ep moe layer to gracefully handle expert_ids that do not belong to local moe rank + for expert_id in range(num_experts): + curr_expert_weight = loaded_weight[expert_id] + weight_loader( + param, + curr_expert_weight, + name, + shard_id, + expert_id, + ) + return True + + loaded_params: Set[str] = set() + params_dict = dict(self.named_parameters(remove_duplicate=False)) + + for name, loaded_weight in weights: + if "rotary_emb.inv_freq" in name: + continue + if "mtp" in name: + continue + if "language_model" in name: + name = name.replace(r"model.language_model.", r"model.") + if ".self_attn." in name: + name = name.replace(".self_attn", "") + if ( + self.config.tie_word_embeddings + and self.pp_group.is_last_rank + and "model.embed_tokens.weight" in name + ): + if "lm_head.weight" in params_dict: + lm_head_param = params_dict["lm_head.weight"] + weight_loader = getattr( + lm_head_param, "weight_loader", default_weight_loader + ) + weight_loader(lm_head_param, loaded_weight) + + layer_id = get_layer_id(name) + if ( + layer_id is not None + and hasattr(self, "start_layer") + and (layer_id < self.start_layer or layer_id >= self.end_layer) + ): + continue + + if self.enable_shared_expert_fusion: + if "mlp.shared_expert." in name: + # Firstly map mlp.shared_expert.xx_proj to mlp.experts.512.xx_proj + name = name.replace( + "mlp.shared_expert.", + f"mlp.experts.{num_experts}.", + ) + + for param_name, weight_name, shard_id in stacked_params_mapping: + if name.endswith("experts.gate_up_proj") or name.endswith( + "experts.down_proj" + ): + is_fused_expert = True + expert_params_mapping = fused_expert_params_mapping + + # Skip non-stacked layers and experts (experts handled below). + if weight_name not in name: + continue + if "visual" in name: + continue + + # We have mlp.experts[0].gate_proj in the checkpoint. + # Since we handle the experts below in expert_params_mapping, + # we need to skip here BEFORE we update the name, otherwise + # name will be updated to mlp.experts[0].gate_up_proj, which + # will then be updated below in expert_params_mapping + # for mlp.experts[0].gate_gate_up_proj, which breaks load. + if "mlp.experts" in name: + continue + name = name.replace(weight_name, param_name) + # Skip loading extra parameters for GPTQ/modelopt models. + if name.endswith(ignore_suffixes) and name not in params_dict: + continue + + if name not in params_dict: + continue + + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + # Track if this is an expert weight to enable early skipping + is_expert_weight = False + + for mapping in expert_params_mapping: + param_name, weight_name, expert_id, shard_id = mapping + if weight_name not in name: + continue + if "visual" in name or self.config.encoder_only: + continue + # Anyway, this is an expert weight and should not be + # attempted to load as other weights later + is_expert_weight = True + name_mapped = name.replace(weight_name, param_name) + if is_fused_expert: + # is_fused_expert is True, the checkpoint contains gate_up_proj and down_proj for each expert + if "experts.gate_up_proj" in name: + # experts.gate_up_proj contains all 512 routed experts, excluding shared experts + # split into w1 and w3 + loaded_weight = loaded_weight.chunk(2, dim=-2) + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight[0], + "w1", + num_experts, + ) + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight[1], + "w3", + num_experts, + ) + elif "experts.down_proj" in name: + # experts.down_proj contains all 512 routed experts, excluding shared experts + load_fused_expert_weights( + name_mapped, + params_dict, + loaded_weight, + shard_id, + num_experts, + ) + elif self.enable_shared_expert_fusion: + # shared experts should be loaded to experts.w13_weight and experts.w2_weight + param = params_dict[name_mapped] + weight_loader = getattr( + param, "weight_loader", default_weight_loader + ) + param = params_dict[name_mapped] + if f"{num_experts}.gate_up_proj" in name: + # split into w1 and w3 + loaded_weight = loaded_weight.chunk(2, dim=-2) + # load to experts.w13_weight, shard_id = w1, expert_id = 512 + weight_loader( + param, + loaded_weight[0], + name_mapped, + "w1", + expert_id, + ) + # load to experts.w13_weight, shard_id = w3, expert_id = 512 + weight_loader( + param, + loaded_weight[1], + name_mapped, + "w3", + expert_id, + ) + else: + # load down_proj to experts.w2_weight, shard_id = w2, expert_id = 512 + # Or load gate_proj and up_proj to experts.w13_weight, shard_id = w1/w3, expert_id = 512 + weight_loader( + param, + loaded_weight, + name_mapped, + shard_id, + expert_id, + ) + else: + # Skip loading extra parameters for GPTQ models. + if ( + name_mapped.endswith(ignore_suffixes) + and name_mapped not in params_dict + ): + continue + param = params_dict[name_mapped] + # We should ask the weight loader to return success or + # not here since otherwise we may skip experts with + # # other available replicas. + weight_loader = param.weight_loader + weight_loader( + param, + loaded_weight, + name_mapped, + shard_id=shard_id, + expert_id=expert_id, + ) + name = name_mapped + break + else: + if is_expert_weight: + # This is an expert weight but not mapped to this rank, skip all remaining processing + continue + + if "visual" in name: + # adapt to VisionAttention + name = name.replace(r"attn.qkv.", r"attn.qkv_proj.") + name = name.replace(r"model.visual.", r"visual.") + + # Skip loading extra parameters for GPTQ/modelopt models. + if name.endswith(ignore_suffixes) and name not in params_dict: + continue + + if name in params_dict.keys(): + param = params_dict[name] + weight_loader = getattr( + param, "weight_loader", default_weight_loader + ) + weight_loader(param, loaded_weight) + else: + logger.warning(f"Parameter {name} not found in params_dict") + loaded_params.add(name) + + self._routed_experts_weights_of_layer = LazyValue( + lambda: { + layer_id: self.model.layers[layer_id].mlp.get_moe_weights() + for layer_id in range(self.model.start_layer, self.model.end_layer) + if isinstance(self.model.layers[layer_id].mlp, Qwen2MoeSparseMoeBlock) + } + ) + + return loaded_params + + @property + def routed_experts_weights_of_layer(self): + return self._routed_experts_weights_of_layer.value + + @classmethod + def get_model_config_for_expert_location(cls, config): + text_config = getattr(config, "text_config", config) + return ModelConfigForExpertLocation( + num_layers=text_config.num_hidden_layers, + num_logical_experts=text_config.num_experts, + num_groups=None, + ) + + +def _qwen3_5_shared_experts_fusion_disable_reason(hf_config, quant_config): + """Why this Qwen3.5 checkpoint cannot fuse its shared expert, or None. + + ROCm-only: an MXFP4 checkpoint cannot fuse, and the model still wants the + #25885 multi-streaming path. Asked by the loader before any layer is built, + so it resolves the text config itself -- the loader hands over whichever + config the entry class takes. + """ + if not _is_hip: + return None + text_config = getattr(hf_config, "text_config", hf_config) + if getattr(text_config, "model_type", None) != "qwen3_5_moe_text": + return None + if can_fuse_shared_expert(text_config, quant_config): + return None + return ( + "Qwen3.5: shared-expert fusion not supported for this checkpoint " + "(multi-streaming #25885 still applies)." + ) + + +# Every class the loader may instantiate for a Qwen3.5 checkpoint answers the +# fusion question the same way. +for _entry_class in ( + Qwen3_5ForCausalLM, + Qwen3_5MoeForCausalLM, + Qwen3_5ForConditionalGeneration, + Qwen3_5MoeForConditionalGeneration, +): + _entry_class.shared_experts_fusion_disable_reason = staticmethod( + _qwen3_5_shared_experts_fusion_disable_reason + ) + + +EntryClass = [Qwen3_5MoeForConditionalGeneration, Qwen3_5ForConditionalGeneration]