A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 17:30 UTC
[BugFix] Fix DeepSeek-R1 hang with DP and MTP (#30119)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com> Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com> Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Lucas Wilkinson committed
Dec 9, 2025 at 13:51 UTC
95501a70ec69d182b124774ff708c3050ab4e91e
1 file changed
+12
-3
vllm/v1/worker/gpu_model_runner.py
+12
-3
@@ -4168,10 +4168,19 @@ class GPUModelRunner(
4168
4169
if self.speculative_config and self.speculative_config.use_eagle():
4170
assert isinstance(self.drafter, EagleProposer)
4171
+ # Eagle currently only supports PIECEWISE cudagraphs.
4172
+ # Therefore only use cudagraphs if the main model uses PIECEWISE
4173
+ # NOTE(lucas): this is a hack, need to clean up.
4174
use_cudagraphs = (
4172
- cudagraph_runtime_mode.has_mode(CUDAGraphMode.PIECEWISE)
4173
- and not self.speculative_config.enforce_eager
4174
- )
4175
+ (
4176
+ is_graph_capturing
4177
+ and cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
4178
+ )
4179
+ or (
4180
+ not is_graph_capturing
4181
+ and cudagraph_runtime_mode != CUDAGraphMode.NONE
4182
+ )
4183
+ ) and not self.speculative_config.enforce_eager
4184
4185
# Note(gnovack) - We need to disable cudagraphs for one of the two
4186
# lora cases when cudagraph_specialize_lora is enabled. This is a