xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-03 07:24:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

updated

Browse Source 
Signed-off-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
Robert Shaw 2025-03-28 01:54:01 +00:00
parent b313220727
commit c5d963835b
2 changed files with 16 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/envs.py
View File

@ -104,6 +104,7 @@ if TYPE_CHECKING:
VLLM_V0_USE_OUTLINES_CACHE: bool = False VLLM_V0_USE_OUTLINES_CACHE: bool = False
VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION: bool = False VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION: bool = False
VLLM_TPU_BUCKET_PADDING_GAP: int = 0 VLLM_TPU_BUCKET_PADDING_GAP: int = 0
VLLM_TPU_DISABLE_SAMPLER_DEBUG: bool = False
def get_default_cache_root(): def get_default_cache_root():
@ -673,6 +674,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_TPU_BUCKET_PADDING_GAP": "VLLM_TPU_BUCKET_PADDING_GAP":
lambda: int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"]) lambda: int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"])
if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0, if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
# Disable sampler path for debugging performance.
"VLLM_TPU_DISABLE_SAMPLER_DEBUG":
lambda: os.environ.get("VLLM_TPU_DISABLE_SAMPLER_DEBUG", "0") == "1",
} }
# end-env-vars-definition # end-env-vars-definition

15
vllm/v1/worker/tpu_model_runner.py
View File

@ -612,10 +612,17 @@ class TPUModelRunner:
kv_caches=self.kv_caches, kv_caches=self.kv_caches,
inputs_embeds=inputs_embeds, inputs_embeds=inputs_embeds,
) )
selected_token_ids = self.model.sample_from_hidden(
hidden_states, tpu_sampling_metadata) if envs.VLLM_TPU_DISABLE_SAMPLER_DEBUG:
# Remove padding on cpu and keep dynamic op outside of xla graph. selected_token_ids = self.model.compute_logits(hidden_states,
selected_token_ids = selected_token_ids.cpu()[:num_reqs] logits_indices, None)
selected_token_ids = selected_token_ids.cpu()[:num_reqs]
else:
selected_token_ids = self.model.sample_from_hidden(
hidden_states, tpu_sampling_metadata)
# Remove padding on cpu and keep dynamic op outside of xla graph.
selected_token_ids = selected_token_ids.cpu()[:num_reqs]
# Update the cache state concurrently. Code above will not block until # Update the cache state concurrently. Code above will not block until
# we use `selected_token_ids`. Add mark_step if post-processing changes # we use `selected_token_ids`. Add mark_step if post-processing changes