[compile] Enable sequence parallelism for full cuda graph without specifying compile sizes (#26681)

Signed-off-by: angelayi <yiangela7@gmail.com>

vllm/compilation/collective_fusion.py

@@ -431,8 +431,15 @@ class AsyncTPPass(VllmPatternMatcherPass):
 
         self.dump_patterns(config, self.patterns)
 
-    def is_applicable_for_shape(self, shape: int | None) -> bool:
+    def is_applicable(self, shape: int | None) -> bool:
-        # only do replace for specific shapes
+        # This pass is applied on top of the sequence parallelism pass.
+        # It inherits the same applicability condition as `SequenceParallelismPass`.
+        # See `SequenceParallelismPass.is_applicable` for more details.
+        if (
+            not self.compilation_config.splitting_ops
+            or self.compilation_config.use_inductor_graph_partition
+        ):
+            return True
         tp_size = get_tensor_model_parallel_world_size()
         return shape is not None and shape % tp_size == 0
 

vllm/compilation/inductor_pass.py

@@ -96,7 +96,7 @@ class InductorPass(CustomGraphPass):
         encoded = json.dumps(dict_, sort_keys=True).encode("utf-8")
         return hashlib.sha256(encoded).hexdigest()
 
-    def is_applicable_for_shape(self, shape: int | None):
+    def is_applicable(self, shape: int | None):
         return True
 
 

vllm/compilation/pass_manager.py

@@ -71,9 +71,11 @@ class PostGradPassManager(CustomGraphPass):
 
         shape = get_pass_context().runtime_shape
         for pass_ in self.passes:
-            if pass_.is_applicable_for_shape(shape):
+            if pass_.is_applicable(shape):
                 pass_(graph)
                 VllmInductorPass.dump_prefix += 1
+            else:
+                logger.debug("Skipping %s with shape %s", pass_, shape)
 
         # post-cleanup goes before fix_functionalization
         # because it requires a functional graph

vllm/compilation/sequence_parallelism.py

@@ -482,7 +482,25 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
             ).register(self.patterns)
         self.dump_patterns(config, self.patterns)
 
-    def is_applicable_for_shape(self, shape: int | None) -> bool:
+    def is_applicable(self, shape: int | None) -> bool:
+        # When sequence parallelism is enabled, the residual tensor from RMSNorm
+        # needs to be split along the sequence dimension. However, this dimension
+        # is symbolic during piecewise compilation, and splitting symbolic shapes
+        # is not supported.
+        #
+        # This pass is therefore only applied when the sequence dimension is
+        # concrete:
+        # 1. In full-graph compilation mode (no Dynamo splitting ops are used).
+        #   For this case we always pad num_tokens to be a multiple of
+        #   tensor_parallel_size, so there's no need to check shape % tp_size == 0.
+        # 2. For specific shape provided during compilation (e.g., from
+        #    `compile_sizes`), which must be divisible by the tensor-parallel
+        #    size.
+        if (
+            not self.compilation_config.splitting_ops
+            or self.compilation_config.use_inductor_graph_partition
+        ):
+            return True
         tp_size = get_tensor_model_parallel_world_size()
         return shape is not None and shape % tp_size == 0
 

vllm/compilation/vllm_inductor_pass.py

@@ -3,6 +3,7 @@
 import functools
 import operator
 import time
+import weakref
 from typing import ClassVar
 
 import regex as re
@@ -28,6 +29,7 @@ class VllmInductorPass(InductorPass):
     """Keep track of pass index for debug dump ordering."""
 
     def __init__(self, config: VllmConfig):
+        self.compilation_config = weakref.proxy(config.compilation_config)
         self.pass_config = config.compilation_config.pass_config
         self.model_dtype = config.model_config.dtype if config.model_config else None
         self.device = config.device_config.device if config.device_config else None