xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 00:45:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

debug

Browse Source 
Signed-off-by: Robert Shaw <rshaw@neuralmagic.com>
This commit is contained in:
Robert Shaw 2025-10-03 13:25:00 -04:00
parent d76541a6c5
commit 52a7d91980
2 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/v1/worker/gpu_model_runner.py
View File

@ -2400,6 +2400,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
# Run the model. # Run the model.
# Use persistent buffers for CUDA graphs. # Use persistent buffers for CUDA graphs.
logger.info(f"====== EXECUTE {ubatch_slices=}, {num_input_tokens=}, {num_tokens_across_dp=}")
with (set_forward_context( with (set_forward_context(
attn_metadata, attn_metadata,
self.vllm_config, self.vllm_config,
@ -3046,6 +3047,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
(1 token) and prefill (multiple tokens) requests. (1 token) and prefill (multiple tokens) requests.
remove_lora: If False, dummy LoRAs are not destroyed after the run remove_lora: If False, dummy LoRAs are not destroyed after the run
""" """
logger.info("====== DUMMY RUN")
assert cudagraph_runtime_mode is None or \ assert cudagraph_runtime_mode is None or \
cudagraph_runtime_mode.valid_runtime_modes() cudagraph_runtime_mode.valid_runtime_modes()

2
vllm/v1/worker/ubatch_splitting.py
View File

@ -167,6 +167,7 @@ def ubatch_split(
num_tokens_unpadded, num_tokens_unpadded,
uniform_decode=uniform_decode, uniform_decode=uniform_decode,
) )
logger.info(f"==== {should_attempt_ubatching=}, {num_tokens_unpadded=}")
# Don't microbatch unless every other DP worker is also microbatching # Don't microbatch unless every other DP worker is also microbatching
should_ubatch, num_tokens_after_padding = get_dp_padding_ubatch( should_ubatch, num_tokens_after_padding = get_dp_padding_ubatch(
@ -175,6 +176,7 @@ def ubatch_split(
should_attempt_ubatching, should_attempt_ubatching,
vllm_config, vllm_config,
) )
logger.info(f"==== {should_ubatch=}, {num_tokens_after_padding=}")
if not should_ubatch: if not should_ubatch:
return (None, None) return (None, None)