[Bugfix][Core] Reserve the KV null block when validating max_model_len (#47272)
Signed-off-by: hyungjun oh <92hyungjun@gmail.com> Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
co-authored by
Claude
Nick Hill
parent
c0233fcf01
commit
76fb6d210a
@@ -77,8 +77,12 @@ def test_auto_fit_max_model_len_rejects_oversized_input(
|
||||
|
||||
# Use a small KV cache budget to force auto-fit to a small
|
||||
# max_model_len. Pin block_size=16 so the budget is independent
|
||||
# of the platform's default block size.
|
||||
kv_cache_bytes = 1_000_000 # 1 MB
|
||||
# of the platform's default block size. One block for this model is
|
||||
# 2 (K/V) * 16 (block) * 12 (heads) * 64 (head_dim) * 2 (fp16)
|
||||
# * 12 (layers) = 589,824 bytes, and the pool must cover at least
|
||||
# two: one is reserved as the null block, so a 1 MB single-block
|
||||
# pool cannot serve any tokens at all.
|
||||
kv_cache_bytes = 2_000_000 # 2 MB = 3 blocks -> 2 usable
|
||||
|
||||
with vllm_runner(
|
||||
model_name=model,
|
||||
|
||||
Reference in New Issue
Block a user