[BugFix] fix some typos found by typos. (#16314)
Signed-off-by: yihong0618 <zouzou0208@gmail.com>
This commit is contained in:
@@ -326,7 +326,7 @@ class FlashAttentionMetadata(AttentionMetadata):
|
||||
assert self.use_cuda_graph
|
||||
|
||||
if turn_prefills_into_decodes:
|
||||
# When Mutli-Step is enabled with Chunked-Prefill, prefills and
|
||||
# When Multi-Step is enabled with Chunked-Prefill, prefills and
|
||||
# decodes are scheduled together. In the first step, all the
|
||||
# prefills turn into decodes. This update reflects that
|
||||
# conversion.
|
||||
|
||||
@@ -152,11 +152,11 @@ class HPUAttentionImpl(AttentionImpl, torch.nn.Module):
|
||||
logger.warning("Could not import HPU FusedSDPA kernel. "
|
||||
"vLLM will use native implementation.")
|
||||
|
||||
suppored_head_sizes = HPUPagedAttention.get_supported_head_sizes()
|
||||
if head_size not in suppored_head_sizes:
|
||||
supported_head_sizes = HPUPagedAttention.get_supported_head_sizes()
|
||||
if head_size not in supported_head_sizes:
|
||||
raise ValueError(
|
||||
f"Head size {head_size} is not supported by PagedAttention. "
|
||||
f"Supported head sizes are: {suppored_head_sizes}.")
|
||||
f"Supported head sizes are: {supported_head_sizes}.")
|
||||
|
||||
if attn_type != AttentionType.DECODER:
|
||||
raise NotImplementedError("Encoder self-attention and "
|
||||
|
||||
@@ -83,8 +83,8 @@ spda_o = scaled_dot_product_attention(
|
||||
return spda_o @ W_O
|
||||
|
||||
NOTE: in the actual code,
|
||||
`kv_b_proj` is [W_UK; W_UV] concatnated per head
|
||||
`q_b_proj` is [W_UQ; W_QR] concatnated per head
|
||||
`kv_b_proj` is [W_UK; W_UV] concatenated per head
|
||||
`q_b_proj` is [W_UQ; W_QR] concatenated per head
|
||||
`out_proj` is W_O
|
||||
|
||||
|
||||
@@ -667,7 +667,7 @@ class MLACommonMetadata(AttentionMetadata):
|
||||
assert num_seqs > num_queries
|
||||
|
||||
if turn_prefills_into_decodes:
|
||||
# When Mutli-Step is enabled with Chunked-Prefill, prefills and
|
||||
# When Multi-Step is enabled with Chunked-Prefill, prefills and
|
||||
# decodes are scheduled together. In the first step, all the
|
||||
# prefills turn into decodes. This update reflects that
|
||||
# conversion.
|
||||
|
||||
@@ -414,11 +414,11 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
|
||||
assert self.num_heads % self.num_kv_heads == 0
|
||||
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
|
||||
|
||||
suppored_head_sizes = PagedAttention.get_supported_head_sizes()
|
||||
if head_size not in suppored_head_sizes:
|
||||
supported_head_sizes = PagedAttention.get_supported_head_sizes()
|
||||
if head_size not in supported_head_sizes:
|
||||
raise ValueError(
|
||||
f"Head size {head_size} is not supported by PagedAttention. "
|
||||
f"Supported head sizes are: {suppored_head_sizes}.")
|
||||
f"Supported head sizes are: {supported_head_sizes}.")
|
||||
|
||||
self.attn_type = attn_type
|
||||
|
||||
|
||||
@@ -446,7 +446,7 @@ def flash_paged_attention(
|
||||
IO tensor dtypes:
|
||||
- This kernel assumes all IO tensors have the same dtype except for
|
||||
block_tables (int32) and mask (int32)
|
||||
- If mixed_percision is True, then all Tensor Engine operation will be
|
||||
- If mixed_precision is True, then all Tensor Engine operation will be
|
||||
performed in bfloat16 and accumulation will be performed in float32.
|
||||
Otherwise the intermediates will be in the same type as the inputs.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user