A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 01:30 UTC
[Misc] Fix typo: seperator -> separator in flashmla_sparse.py (#32411)
Signed-off-by: Guofang Tang <tinggofun@gmail.com> Co-authored-by: Guofang Tang <tinggofun@gmail.com>
Guofang.Tang committed
Jan 17, 2026 at 20:18 UTC
2b99f210f53faee71fd56cf52036040847e6b420
1 file changed
+7
-7
vllm/v1/attention/backends/mla/flashmla_sparse.py
+7
-7
@@ -149,7 +149,7 @@ class FlashMLASparseMetadata(AttentionMetadata):
149
cache_lens: torch.Tensor
150
151
@dataclass
152
- class FP8SeperatePrefillDecode:
152
+ class FP8SeparatePrefillDecode:
153
@dataclass
154
class Decode:
155
kernel_metadata: "FlashMLASparseMetadata.FP8KernelMetadata"
@@ -196,7 +196,7 @@ class FlashMLASparseMetadata(AttentionMetadata):
196
decode: Decode | None = None
197
prefill: Prefill | None = None
198
199
- fp8_extra_metadata: FP8SeperatePrefillDecode | FP8KernelMetadata | None = None
199
+ fp8_extra_metadata: FP8SeparatePrefillDecode | FP8KernelMetadata | None = None
200
fp8_use_mixed_batch: bool = False
201
202
@@ -485,7 +485,7 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
485
def _build_fp8_separate_prefill_decode(
486
self,
487
common_attn_metadata: CommonAttentionMetadata,
488
- ) -> "FlashMLASparseMetadata.FP8SeperatePrefillDecode":
488
+ ) -> "FlashMLASparseMetadata.FP8SeparatePrefillDecode":
489
num_tokens = common_attn_metadata.num_actual_tokens
490
491
(num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens) = (
@@ -496,7 +496,7 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
496
)
497
)
498
499
- FP8Meta = FlashMLASparseMetadata.FP8SeperatePrefillDecode
499
+ FP8Meta = FlashMLASparseMetadata.FP8SeparatePrefillDecode
500
fp8_metadata = FP8Meta(
501
num_decodes=num_decodes,
502
num_prefills=num_prefills,
@@ -659,7 +659,7 @@ class FlashMLASparseMetadataBuilder(AttentionMetadataBuilder[FlashMLASparseMetad
659
req_id_per_token = self.req_id_per_token_buffer[:num_tokens]
660
661
fp8_extra_metadata: (
662
- FlashMLASparseMetadata.FP8SeperatePrefillDecode
662
+ FlashMLASparseMetadata.FP8SeparatePrefillDecode
663
| FlashMLASparseMetadata.FP8KernelMetadata
664
| None
665
) = None
@@ -765,7 +765,7 @@ class FlashMLASparseImpl(MLACommonBaseImpl[FlashMLASparseMetadata]):
765
attn_metadata: FlashMLASparseMetadata,
766
) -> torch.Tensor:
767
fp8_metadata = attn_metadata.fp8_extra_metadata
768
- assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeperatePrefillDecode)
768
+ assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeparatePrefillDecode)
769
num_decodes = fp8_metadata.num_decodes
770
771
prefill_request_ids = None
@@ -794,7 +794,7 @@ class FlashMLASparseImpl(MLACommonBaseImpl[FlashMLASparseMetadata]):
794
)
795
796
fp8_metadata = attn_metadata.fp8_extra_metadata
797
- assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeperatePrefillDecode)
797
+ assert isinstance(fp8_metadata, FlashMLASparseMetadata.FP8SeparatePrefillDecode)
798
799
def _fp8_decode(q: torch.Tensor, topk_indices: torch.Tensor) -> torch.Tensor:
800
# Reshape q: (num_decode_tokens, num_heads, head_dim)