From 24fab45d96a91b491db338ee02cd24e55b7fbb5f Mon Sep 17 00:00:00 2001
From: Michael Goin <mgoin64@gmail.com>
Date: Tue, 23 Sep 2025 15:29:26 -0400
Subject: [PATCH] [Perf] Change default CUDAGraphMode from PIECEWISE to
 FULL_AND_PIECEWISE (#25444)

Signed-off-by: mgoin <mgoin64@gmail.com>
---
 vllm/config/__init__.py            |  9 ++++++++-
 vllm/config/compilation.py         |  7 +++----
 vllm/v1/worker/gpu_model_runner.py | 23 +++++++++++++++++++++--
 3 files changed, 32 insertions(+), 7 deletions(-)

diff --git a/vllm/config/__init__.py b/vllm/config/__init__.py
index 50e8cad23617..d786d3e289b3 100644
--- a/vllm/config/__init__.py
+++ b/vllm/config/__init__.py
@@ -509,8 +509,15 @@ class VllmConfig:
             if self.compilation_config.cudagraph_mode is None:
                 if envs.VLLM_USE_V1 and self.compilation_config.level \
                     == CompilationLevel.PIECEWISE:
+                    # default to full and piecewise for most models
                     self.compilation_config.cudagraph_mode = \
-                        CUDAGraphMode.PIECEWISE
+                        CUDAGraphMode.FULL_AND_PIECEWISE
+
+                    # pooling model does not support full cudagraphs
+                    if self.model_config is not None and \
+                        self.model_config.pooler_config is not None:
+                        self.compilation_config.cudagraph_mode = \
+                            CUDAGraphMode.PIECEWISE
                 else:
                     self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE
 
diff --git a/vllm/config/compilation.py b/vllm/config/compilation.py
index 34fa7fcfe7e8..0441745e8b36 100644
--- a/vllm/config/compilation.py
+++ b/vllm/config/compilation.py
@@ -228,15 +228,14 @@ class CompilationConfig:
     The mode of the cudagraph:
 
     - NONE, no cudagraph capture.
-    - PIECEWISE. (v1 default)
+    - PIECEWISE.
     - FULL.
     - FULL_DECODE_ONLY.
-    - FULL_AND_PIECEWISE.
+    - FULL_AND_PIECEWISE. (v1 default)
 
     PIECEWISE mode build piecewise cudagraph only, keeping the cudagraph
     incompatible ops (i.e. some attention ops) outside the cudagraph
     for general flexibility.
-    This is the default mode.
 
     FULL mode: Capture full cudagraph for all batches. Can be good for small
     models or workloads with small prompts; not supported by many backends.
@@ -249,7 +248,7 @@ class CompilationConfig:
     
     FULL_AND_PIECEWISE mode: Capture full cudagraph for decode batches and
     piecewise cudagraph for prefill and mixed prefill-decode batches.
-    This is like the most performant mode for most models.
+    This is the most performant mode for most models and is the default.
 
     Currently, the cudagraph mode is only used for the v1 engine.
     Note that the cudagraph logic is generally orthogonal to the 
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index ed324138c6fe..df4f76958586 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -2947,8 +2947,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                 # TODO(luka) better system for describing dummy batches
                 seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
             else:
-                # Make sure max_model_len is used at the graph capture time.
-                seq_lens = self.max_model_len
+                seq_lens = max_query_len
             self.seq_lens.np[:num_reqs] = seq_lens
             self.seq_lens.np[num_reqs:] = 0
             self.seq_lens.copy_to_gpu()
@@ -3541,6 +3540,26 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
                     CUDAGraphMode.FULL_DECODE_ONLY
             logger.warning(msg)
 
+        # check that if we are doing decode full-cudagraphs it is supported
+        if (cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
+                and min_cg_support == AttentionCGSupport.NEVER):
+            msg = (f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
+                   f"with {min_cg_builder_name} backend (support: "
+                   f"{min_cg_support})")
+            if (self.compilation_config.level == CompilationLevel.PIECEWISE and
+                (self.compilation_config.splitting_ops_contain_attention()
+                 or self.compilation_config.use_inductor_graph_partition)):
+                msg += "; setting cudagraph_mode=PIECEWISE because "\
+                    "attention is compiled piecewise"
+                cudagraph_mode = self.compilation_config.cudagraph_mode = \
+                    CUDAGraphMode.PIECEWISE
+            else:
+                msg += "; setting cudagraph_mode=NONE because "\
+                    "attention is not compiled piecewise"
+                cudagraph_mode = self.compilation_config.cudagraph_mode = \
+                    CUDAGraphMode.NONE
+            logger.warning(msg)
+
         # check that if we are doing spec-decode + decode full-cudagraphs it is
         # supported
         if (cudagraph_mode.decode_mode() == CUDAGraphMode.FULL