[None][chroe] Polish qwen3-next modeling code.

Signed-off-by: nv-guomingz <[email protected]>

Author: nv-guomingz

tensorrt_llm/_torch/models/modeling_qwen3_next.py

@@ -50,9 +50,10 @@
 from ..modules.linear import Linear, TensorParallelMode
 from ..modules.mamba.causal_conv1d import causal_conv1d_fn, causal_conv1d_update
 from ..modules.mamba.layernorm_gated import RMSNorm as RMSNormGated
+from ..modules.multi_stream_utils import maybe_execute_in_parallel
 from ..modules.rms_norm import RMSNorm
 from ..speculative import SpecMetadata
-from ..utils import AuxStreamType
+from ..utils import AuxStreamType, EventType
 from .modeling_qwen3 import Qwen3Attention
 from .modeling_speculative import SpecDecOneEngineForCausalLM
 from .modeling_utils import DecoderModel, EagerFusionConfig, register_auto_model
@@ -387,6 +388,7 @@ def __init__(
         self.mapping = model_config.mapping
         self.allreduce = AllReduce(mapping=model_config.mapping,
                                    strategy=model_config.allreduce_strategy)
+        self.aux_stream = aux_stream
 
         self.gate = Qwen3NextGate(
             hidden_size=self.hidden_dim,
@@ -425,6 +427,11 @@ def __init__(
                                          dtype=config.torch_dtype,
                                          quant_config=None)
 
+        self.event_dict = {
+            key: torch.cuda.Event()
+            for key in [EventType.Main, EventType.MoeShared]
+        }
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -450,22 +457,33 @@ def forward(
                                           dim=0,
                                           sizes=all_rank_num_tokens)
 
-        router_logits = self.gate(hidden_states)
-        final_hidden_states = self.experts(
-            hidden_states,
-            router_logits,
-            all_rank_num_tokens=all_rank_num_tokens,
-            use_dp_padding=use_dp_padding,
-            do_finalize=do_finalize,
-        )
+        def _compute_routed_output():
+            router_logits = self.gate(hidden_states)
+            final_hidden_states = self.experts(
+                hidden_states,
+                router_logits,
+                all_rank_num_tokens=all_rank_num_tokens,
+                use_dp_padding=use_dp_padding,
+                do_finalize=do_finalize,
+            )
+            return final_hidden_states
 
+        def _compute_shared_output():
+            shared_expert_output = self.shared_expert(hidden_states)
+            shared_expert_output = F.sigmoid(
+                self.shared_expert_gate(hidden_states)) * shared_expert_output
+            return shared_expert_output
+
+        final_hidden_states, shared_expert_output = maybe_execute_in_parallel(
+            _compute_routed_output,
+            _compute_shared_output,
+            self.event_dict[EventType.Main],
+            self.event_dict[EventType.MoeShared],
+            self.aux_stream,
+        )
         if not do_finalize:
             return final_hidden_states
 
-        shared_expert_output = self.shared_expert(hidden_states)
-        shared_expert_output = F.sigmoid(
-            self.shared_expert_gate(hidden_states)) * shared_expert_output
-
         final_hidden_states = final_hidden_states + shared_expert_output
 
         if not self.enable_attention_dp and self.mapping.tp_size > 1:
@@ -1038,8 +1056,6 @@ def forward(
                 self.head_v_dim,
             )
         else:
-            query, key, value, z, b, a = self.fix_query_key_value_ordering(
-                projected_states_qkvz, projected_states_ba)
             query, key, value = map(lambda x: x.reshape(x.shape[0], -1),
                                     (query, key, value))
             mixed_qkv = torch.cat((query, key, value), dim=-1)
@@ -1061,16 +1077,11 @@ def forward(
             "num_decode": num_decodes,
         }
 
-        new_implementation = True
-        if new_implementation:
-            if num_prefills > 0:
-                attn_out = self.forward_extend(conv_states, ssm_states,
-                                               **kwargs)
-            else:
-                attn_out = self.forward_decode(conv_states, ssm_states,
-                                               num_decodes,
-                                               mamba_metadata.cu_seqlens,
-                                               **kwargs)
+        if num_prefills > 0:
+            attn_out = self.forward_extend(conv_states, ssm_states, **kwargs)
+        else:
+            attn_out = self.forward_decode(conv_states, ssm_states, num_decodes,
+                                           mamba_metadata.cu_seqlens, **kwargs)
 
         z_shape_og = z.shape
         # reshape input data into 2D tensor
@@ -1125,7 +1136,7 @@ def __init__(
             "TRTLLM_QWEN3_EAGER_FUSION_DISABLED", "1") == "0"
         self.enable_fusion &= not self.enable_attention_dp
 
-        self.mapping.has_tp()
+        # has_tp = self.mapping.has_tp()
         has_pp = self.mapping.has_pp()
 
         # self.fusion_config.PRE_MOE_FUSION = self.enable_fusion and has_tp
@@ -1284,7 +1295,7 @@ def __init__(self, model_config: ModelConfig[Qwen3NextConfig],
             "TRTLLM_QWEN3_EAGER_FUSION_DISABLED", "0") == "0"
         self.enable_fusion &= not self.enable_attention_dp
 
-        self.mapping.has_tp()
+        # has_tp = self.mapping.has_tp()
         has_pp = self.mapping.has_pp()
 
         # self.fusion_config.PRE_MOE_FUSION = self.enable_fusion and has_tp

tensorrt_llm/_torch/modules/fla/chunk.py

@@ -90,8 +90,6 @@ def forward(
         cu_seqlens: Optional[torch.LongTensor] = None,
         use_qk_l2norm_in_kernel: bool = False,
     ):
-        pass
-
         if use_qk_l2norm_in_kernel:
             q = l2norm_fwd(q)
             k = l2norm_fwd(k)