xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-08 14:07:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2025-09-18 18:37:30 -07:00
parent af65838d1f
commit 8af87986aa
2 changed files with 2246 additions and 152 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

@ -307,6 +307,7 @@ class GPUModelRunner:
# Slot mappings: [num_kv_cache_groups, num_tokens]
slot_mappings = self.block_tables.compute_slot_mappings(
query_start_loc_gpu, positions.gpu[:num_tokens])
logits_indices = query_start_loc_gpu[1:] - 1
num_logits_indices = logits_indices.size(0)
@ -366,6 +367,7 @@ class GPUModelRunner:
# TODO(woosuk): Support DP sampler + CUDA graphs.
sample_hidden_states = hidden_states[input_batch.logits_indices]
logits = self.model.compute_logits(sample_hidden_states, None)
pos = input_batch.positions[input_batch.logits_indices]
sampling_metadata = self.req_states.make_sampling_metadata(
input_batch.idx_mapping_np, pos)

2396
vllm/v1/worker/gpu_model_runner.py
View File

File diff suppressed because it is too large Load Diff