[Fix] deepseek_v4: Fix mHC prewarm when chunked prefill is disabled (#43448)
This commit is contained in:
@@ -858,7 +858,11 @@ def prewarm_mhc_pre(
|
||||
from sglang.srt.runtime_context import get_schedule
|
||||
|
||||
hc_mult, hidden_size = residual.shape[-2], residual.shape[-1]
|
||||
max_num_tokens = get_schedule().chunked_prefill_size
|
||||
schedule = get_schedule()
|
||||
max_num_tokens = schedule.chunked_prefill_size
|
||||
if max_num_tokens is None or max_num_tokens <= 0:
|
||||
# Chunked prefill disabled: bound by the prefill token budget.
|
||||
max_num_tokens = schedule.max_prefill_tokens
|
||||
buckets = get_mhc_pre_token_count_representatives(
|
||||
max_num_tokens, hc_mult * hidden_size
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user