[PP] Support prefill CUDA graph proxy tensors (#36248)

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

python/sglang/srt/model_executor/cuda_graph_buffer_registry.py

@@ -953,7 +953,10 @@ def build_prefill_registry(
             def _pp_source(key):
                 def _fn(_fb, ctx):
                     ppx = ctx.pp_proxy_tensors
-                    return None if ppx is None else ppx.tensors[key]
+                    # Proxy contracts vary by model. The capture buffers are a
+                    # stable-address superset; only copy fields present in the
+                    # live proxy for this model.
+                    return None if ppx is None else ppx.tensors.get(key)
 
                 return _fn
 

python/sglang/srt/model_executor/model_runner_components/cuda_graph_setup.py

@@ -45,6 +45,7 @@ from sglang.srt.runtime_context import (
     get_disagg,
     get_exec,
     get_flags,
+    get_parallel,
     get_schedule,
     get_spec,
 )
@@ -72,6 +73,8 @@ def _align_pipeline_layers(layers: list, layer_model) -> list:
         f"invalid pipeline layer range [{start_layer}, {end_layer}) for "
         f"{len(layer_model.layers)} layers"
     )
+    if len(layers) == len(layer_model.layers):
+        return layers
     assert (
         len(layers) <= end_layer - start_layer
     ), f"found {len(layers)} layers in PP range [{start_layer}, {end_layer})"
@@ -365,6 +368,17 @@ def capture_prefill_graph(
 
     prefill_config = get_exec().graph.cuda_graph_config.prefill
     prefill_backend = prefill_config.backend
+    parallel = get_parallel()
+    if (
+        prefill_backend == Backend.BREAKABLE
+        and parallel.enable_prefill_cp
+        and parallel.pp_size > 1
+    ):
+        logger.warning(
+            "Disable prefill CUDA graph because pipeline parallelism combined "
+            "with prefill context parallelism is not validated."
+        )
+        return result(eager_runner)
     context_length = model_runner.model_config.context_len
     if prefill_backend == Backend.FULL:
         max_capture_requests = prefill_config.full_prefill_max_req

python/sglang/srt/model_executor/model_runner_components/layer_setup.py

@@ -62,12 +62,11 @@ def compute_attention_and_moe_layers(layer_model: Any) -> AttentionAndMoeLayers:
                 # Mamba layer with split op support - store the layer itself
                 attn_layer = layer
 
-        if attn_layer is not None:
-            attention_layers.append(attn_layer)
-            mha_companion_layers.append(mha_companion_layer)
-        elif hasattr(layer, "mixer"):
-            attention_layers.append(None)
-            mha_companion_layers.append(None)
+        # Keep these lists aligned with global layer ids. Pipeline-parallel
+        # models retain placeholders outside the local stage, while real
+        # attention modules use their global layer_id during graph replay.
+        attention_layers.append(attn_layer)
+        mha_companion_layers.append(mha_companion_layer)
 
         moe_block = None
         moe_fusion = None

python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py

@@ -185,6 +185,23 @@ def _resolve_transformer_layer_model(model: torch.nn.Module) -> torch.nn.Module:
     return layer_model
 
 
+def _build_layer_model_forward_kwargs(
+    layer_model: torch.nn.Module,
+    forward_batch: ForwardBatch,
+    pp_proxy_tensors: Optional[PPProxyTensors],
+) -> Dict[str, Any]:
+    """Bind optional transformer inputs by name across model signatures."""
+    parameters = inspect.signature(layer_model.forward).parameters
+    kwargs = {}
+    for embeds_name in ("input_embeds", "inputs_embeds"):
+        if embeds_name in parameters:
+            kwargs[embeds_name] = forward_batch.input_embeds
+            break
+    if pp_proxy_tensors is not None and "pp_proxy_tensors" in parameters:
+        kwargs["pp_proxy_tensors"] = pp_proxy_tensors
+    return kwargs
+
+
 def _slice_output_rows(output: Any, num_tokens: int) -> Any:
     """Slice every tensor leaf in a transformer-body output by token rows.
 
@@ -333,6 +350,7 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
             dtype=self.model_runner.dtype,
             enable_mamba_track=self.mamba_track_enabled,
             pp_size=self.pp_size,
+            is_first_pp_rank=self.model_runner.pp_group.is_first_rank,
             hc_hidden_size=getattr(
                 self.model_runner.model_config, "hc_hidden_size", None
             ),
@@ -653,12 +671,12 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
 
         return forward_batch.positions
 
-    def _static_pp_proxy_tensors(self, num_tokens: int) -> Optional[PPProxyTensors]:
+    def _capture_pp_proxy_tensors(self, num_tokens: int) -> Optional[PPProxyTensors]:
         buffers = self.buffers.pp_proxy_tensors
-        if buffers is None:
+        if buffers is None or self.model_runner.pp_group.is_first_rank:
             return None
         return PPProxyTensors(
-            {key: value[:num_tokens] for key, value in buffers.items()}
+            {name: buffer[:num_tokens] for name, buffer in buffers.items()}
         )
 
     @contextmanager
@@ -717,20 +735,28 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
         set_is_extend_in_batch(False)
 
         with self._prefill_forward_context(forward_batch):
-            pp_kwargs = self.model_runner._pp_kwargs(
-                self._static_pp_proxy_tensors(num_tokens)
-            )
+            pp_proxy_tensors = self._capture_pp_proxy_tensors(num_tokens)
             if self._uses_eager_prefill_tail():
                 # BCG / Full: capture the transformer body only.
                 positions = self._get_layer_model_positions(forward_batch)
+                input_ids = forward_batch.input_ids
+                kwargs = _build_layer_model_forward_kwargs(
+                    self.layer_model, forward_batch, pp_proxy_tensors
+                )
+                if pp_proxy_tensors is not None:
+                    input_ids = None
+                    for embeds_name in ("input_embeds", "inputs_embeds"):
+                        if embeds_name in kwargs:
+                            kwargs[embeds_name] = None
+                            break
                 return self.layer_model.forward(
-                    forward_batch.input_ids,
+                    input_ids,
                     positions,
                     forward_batch,
-                    forward_batch.input_embeds,
-                    **pp_kwargs,
+                    **kwargs,
                 )
             # tc_piecewise: compile/capture the outer model.forward path.
+            pp_kwargs = self.model_runner._pp_kwargs(pp_proxy_tensors)
             return self.model_runner.model.forward(
                 forward_batch.input_ids,
                 forward_batch.positions,
@@ -1765,6 +1791,10 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
         # uses real request metadata instead of padded slots. BCG has no
         # request-slot padding, so static_forward_batch is already the serving batch.
         tail_batch = forward_batch if full_path else static_forward_batch
+        if not full_path:
+            # MTP consumes the target model's live multimodal embeddings in its
+            # eager wrapper before the captured transformer body is replayed.
+            tail_batch.mm_input_embeds = forward_batch.mm_input_embeds
         try:
             with self._prefill_forward_context(
                 static_forward_batch,
@@ -1836,7 +1866,7 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
         if isinstance(output, EmbeddingPoolerOutput):
             return output
         assert isinstance(output, PPProxyTensors)
-        return output[: self.raw_num_tokens]
+        return _slice_output_rows(output, self.raw_num_tokens)
 
     def _validate_capture_hidden_mode(self, forward_batch: ForwardBatch) -> None:
         if self.capture_hidden_mode < forward_batch.capture_hidden_mode:

python/sglang/srt/model_executor/runner_utils/buffers.py

@@ -60,6 +60,38 @@ def _grouped_foreach_copy_(dsts: List[torch.Tensor], srcs: List[torch.Tensor]) -
         foreach_copy(group_dsts, group_srcs)
 
 
+def _allocate_pp_proxy_tensors(
+    *,
+    max_num_tokens: int,
+    max_hidden_tokens: int,
+    hidden_size: int,
+    dtype: torch.dtype,
+    hc_hidden_size: Optional[int] = None,
+    pp_proxy_topk_size: Optional[int] = None,
+    pp_proxy_residual_num_blocks: Optional[int] = None,
+) -> Dict[str, torch.Tensor]:
+    """Allocate the stable buffers consumed by an incoming PP proxy."""
+    is_mhc = hc_hidden_size is not None
+    pp_hidden_size = hc_hidden_size if is_mhc else hidden_size
+    pp_proxy_tensors = {
+        "hidden_states": torch.zeros((max_hidden_tokens, pp_hidden_size), dtype=dtype),
+    }
+    if not is_mhc:
+        # Only Kimi K3 supplies num_blocks: its PP bank is token-major
+        # [T, blocks, H]. Other models use the phase-specific hidden-token bound.
+        residual_shape = (
+            (max_num_tokens, pp_proxy_residual_num_blocks, hidden_size)
+            if pp_proxy_residual_num_blocks is not None
+            else (max_hidden_tokens, hidden_size)
+        )
+        pp_proxy_tensors["residual"] = torch.zeros(residual_shape, dtype=dtype)
+    if pp_proxy_topk_size is not None:
+        pp_proxy_tensors["topk_indices"] = torch.zeros(
+            (max_num_tokens, pp_proxy_topk_size), dtype=torch.int32
+        )
+    return pp_proxy_tensors
+
+
 @dataclass
 class DecodeInputBuffers(ForwardInputBuffers):
     input_ids: torch.Tensor
@@ -129,29 +161,19 @@ class DecodeInputBuffers(ForwardInputBuffers):
                 torch.zeros((max_bs,), dtype=torch.bool) if enable_mamba_track else None
             )
 
-            if pp_size > 1:
-                is_mhc = hc_hidden_size is not None
-                hs = hc_hidden_size if is_mhc else hidden_size
-                pp_proxy_tensors = {
-                    "hidden_states": torch.zeros((max_num_token, hs), dtype=dtype),
-                }
-                if not is_mhc:
-                    # Only Kimi K3 supplies num_blocks: its PP bank is token-major
-                    # [T, blocks, H]. Other models keep the legacy [max_bs, H].
-                    residual_shape = (
-                        (max_num_token, pp_proxy_residual_num_blocks, hidden_size)
-                        if pp_proxy_residual_num_blocks is not None
-                        else (max_num_token, hidden_size)
-                    )
-                    pp_proxy_tensors["residual"] = torch.zeros(
-                        residual_shape, dtype=dtype
-                    )
-                if pp_proxy_topk_size is not None:
-                    pp_proxy_tensors["topk_indices"] = torch.zeros(
-                        (max_num_token, pp_proxy_topk_size), dtype=torch.int32
-                    )
-            else:
-                pp_proxy_tensors = None
+            pp_proxy_tensors = (
+                _allocate_pp_proxy_tensors(
+                    max_num_tokens=max_num_token,
+                    max_hidden_tokens=max_num_token,
+                    hidden_size=hidden_size,
+                    dtype=dtype,
+                    hc_hidden_size=hc_hidden_size,
+                    pp_proxy_topk_size=pp_proxy_topk_size,
+                    pp_proxy_residual_num_blocks=pp_proxy_residual_num_blocks,
+                )
+                if pp_size > 1
+                else None
+            )
 
             if is_encoder_decoder:
                 encoder_lens = torch.full(
@@ -357,6 +379,7 @@ class PrefillInputBuffers(ForwardInputBuffers):
         dtype: torch.dtype,
         enable_mamba_track: bool,
         pp_size: int = 1,
+        is_first_pp_rank: bool = False,
         hc_hidden_size: Optional[int] = None,
         pp_proxy_topk_size: Optional[int] = None,
         pp_proxy_residual_num_blocks: Optional[int] = None,
@@ -387,29 +410,19 @@ class PrefillInputBuffers(ForwardInputBuffers):
                 input_embeds = None
                 mrope_positions = None
 
-            if pp_size > 1:
-                is_mhc = hc_hidden_size is not None
-                pp_hidden_size = hc_hidden_size if is_mhc else hidden_size
-                pp_proxy_tensors = {
-                    "hidden_states": torch.zeros(
-                        (max_num_tokens, pp_hidden_size), dtype=dtype
-                    )
-                }
-                if not is_mhc:
-                    residual_shape = (
-                        (max_num_tokens, pp_proxy_residual_num_blocks, hidden_size)
-                        if pp_proxy_residual_num_blocks is not None
-                        else (max_num_tokens, hidden_size)
-                    )
-                    pp_proxy_tensors["residual"] = torch.zeros(
-                        residual_shape, dtype=dtype
-                    )
-                if pp_proxy_topk_size is not None:
-                    pp_proxy_tensors["topk_indices"] = torch.zeros(
-                        (max_num_tokens, pp_proxy_topk_size), dtype=torch.int32
-                    )
-            else:
-                pp_proxy_tensors = None
+            pp_proxy_tensors = (
+                _allocate_pp_proxy_tensors(
+                    max_num_tokens=max_num_tokens,
+                    max_hidden_tokens=max_num_tokens,
+                    hidden_size=hidden_size,
+                    dtype=dtype,
+                    hc_hidden_size=hc_hidden_size,
+                    pp_proxy_topk_size=pp_proxy_topk_size,
+                    pp_proxy_residual_num_blocks=pp_proxy_residual_num_blocks,
+                )
+                if pp_size > 1 and not is_first_pp_rank
+                else None
+            )
 
         return cls(
             input_ids=input_ids,