[Perf] Disable clean_logits in deepgemm fp8_mqa_logits kernel (#33568)
This commit is contained in:
@@ -242,6 +242,7 @@ def fp8_mqa_logits(
|
||||
weights: torch.Tensor,
|
||||
cu_seqlen_ks: torch.Tensor,
|
||||
cu_seqlen_ke: torch.Tensor,
|
||||
clean_logits: bool,
|
||||
) -> torch.Tensor:
|
||||
"""Compute FP8 MQA logits for a single sequence without KV paging.
|
||||
|
||||
@@ -256,6 +257,7 @@ def fp8_mqa_logits(
|
||||
shape [M], dtype int32.
|
||||
cu_seqlen_ke: End indices (exclusive) for valid K per query position,
|
||||
shape [M], dtype int32.
|
||||
clean_logits: Whether to clean the unfilled logits into `-inf`.
|
||||
|
||||
Returns:
|
||||
Logits tensor of shape [M, N], dtype `torch.float32`.
|
||||
@@ -263,7 +265,9 @@ def fp8_mqa_logits(
|
||||
_lazy_init()
|
||||
if _fp8_mqa_logits_impl is None:
|
||||
return _missing()
|
||||
return _fp8_mqa_logits_impl(q, kv, weights, cu_seqlen_ks, cu_seqlen_ke)
|
||||
return _fp8_mqa_logits_impl(
|
||||
q, kv, weights, cu_seqlen_ks, cu_seqlen_ke, clean_logits=clean_logits
|
||||
)
|
||||
|
||||
|
||||
def get_paged_mqa_logits_metadata(
|
||||
@@ -295,6 +299,7 @@ def fp8_paged_mqa_logits(
|
||||
block_tables: torch.Tensor,
|
||||
schedule_metadata: torch.Tensor,
|
||||
max_model_len: int,
|
||||
clean_logits: bool,
|
||||
) -> torch.Tensor:
|
||||
"""Compute FP8 MQA logits using paged KV-cache.
|
||||
|
||||
@@ -312,6 +317,7 @@ def fp8_paged_mqa_logits(
|
||||
schedule_metadata: Returned by `get_paged_mqa_logits_metadata`;
|
||||
used to distribute work across SMs.
|
||||
max_model_len: Maximum sequence length used to size the logits output.
|
||||
clean_logits: Whether to clean the unfilled logits into `-inf`.
|
||||
|
||||
Returns:
|
||||
Logits tensor of shape [B * next_n, max_model_len], dtype
|
||||
@@ -328,7 +334,7 @@ def fp8_paged_mqa_logits(
|
||||
block_tables,
|
||||
schedule_metadata,
|
||||
max_model_len,
|
||||
clean_logits=True,
|
||||
clean_logits=clean_logits,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user