[CPU] Fix FP8 attention scratchpad sizing (#50194)

Signed-off-by: Li, Tianmu <tianmu.li@intel.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
Tianmu Li
2026-07-29 15:34:37 +08:00
committed by GitHub
co-authored by Codex Li, Jiang <jiang1.li@intel.com>
parent 100d655a23
commit 65a1a16594
5 changed files with 35 additions and 7 deletions
+2
View File
@@ -3814,6 +3814,7 @@ def cpu_attn_get_scheduler_metadata(
isa: str,
enable_kv_split: bool,
dynamic_causal: torch.Tensor | None = None,
kv_cache_dtype: str = "auto",
) -> torch.Tensor:
scheduler_metadata = torch.ops._C.get_scheduler_metadata(
num_reqs,
@@ -3828,6 +3829,7 @@ def cpu_attn_get_scheduler_metadata(
isa,
enable_kv_split,
dynamic_causal,
kv_cache_dtype,
)
return scheduler_metadata
+3 -2
View File
@@ -155,12 +155,12 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]
if self.window_size is None:
self.window_size = -1
self.block_size = vllm_config.cache_config.block_size
kv_cache_dtype_str = vllm_config.cache_config.cache_dtype
self.kv_cache_dtype = vllm_config.cache_config.cache_dtype
self.isa = _get_attn_isa(
self.dtype,
self.block_size,
self.head_dim,
kv_cache_dtype_str,
self.kv_cache_dtype,
)
self.is_cross_attention = isinstance(kv_cache_spec, CrossAttentionSpec)
self.is_encoder_only_attention = isinstance(
@@ -234,6 +234,7 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]
isa=self.isa,
enable_kv_split=envs.VLLM_CPU_ATTN_SPLIT_KV,
dynamic_causal=dynamic_casual,
kv_cache_dtype=self.kv_cache_dtype,
)
attn_metadata = CPUAttentionMetadata(