[CI] update typos config for CI pre-commit and fix some spells (#20919)
Signed-off-by: Peter Pan <Peter.Pan@daocloud.io>
This commit is contained in:
@@ -961,7 +961,7 @@ class DifferentialFlashAttentionImpl(AttentionImpl):
|
||||
"... H (two D) -> ... (H two) D",
|
||||
two=2)
|
||||
|
||||
else: # re-use the kv cache, full attention
|
||||
else: # reuse the kv cache, full attention
|
||||
q = q.view(-1, self.num_heads, self.head_size)
|
||||
q1, q2 = self.split_heads(q)
|
||||
# kv_cache shape is (2, num_blocks, block_size, num_kv_heads, head_size) # noqa: E501
|
||||
|
||||
@@ -372,7 +372,7 @@ class OpenAIServingResponses(OpenAIServing):
|
||||
})
|
||||
|
||||
# Append the new input.
|
||||
# Reponses API supports simple text inputs without chat format.
|
||||
# Responses API supports simple text inputs without chat format.
|
||||
if isinstance(request.input, str):
|
||||
messages.append({"role": "user", "content": request.input})
|
||||
else:
|
||||
|
||||
@@ -1172,7 +1172,7 @@ def fused_experts(
|
||||
allow_cutlass_block_scaled_grouped_gemm: bool = False) -> torch.Tensor:
|
||||
# For now, disable DeepGemm for small N (<= 512) until better
|
||||
# permute/unpermute ops are available.
|
||||
# However, on B200, we use DeepGemm for all cases becuase they only support
|
||||
# However, on B200, we use DeepGemm for all cases because they only support
|
||||
# E8M0 scale, which means we requantize the weight and input to the specific
|
||||
# scale. Fallen back to cutlass or triton for some cases would cause
|
||||
# accuracy issue.
|
||||
|
||||
@@ -193,7 +193,7 @@ class SambaYAttention(nn.Module):
|
||||
],
|
||||
dim=-1)
|
||||
attn_output = self.attn(q, k, v)
|
||||
else: # re-use the kv cache, full attention
|
||||
else: # reuse the kv cache, full attention
|
||||
q = self.Wqkv(hidden_states)
|
||||
attn_output = self.attn(q, None, None)
|
||||
attn_output = attn_output.view(-1, self.num_heads * self.head_dim)
|
||||
|
||||
@@ -394,7 +394,7 @@ def use_cudnn_prefill() -> bool:
|
||||
|
||||
|
||||
# Currently 394MB, this can be tuned based on GEMM sizes used.
|
||||
# Choosen to be the same as sglang:
|
||||
# Chosen to be the same as sglang:
|
||||
# https://github.com/sgl-project/sglang/blob/766392c6bda2558b61ce6d1c1bfd8081a549e1f1/python/sglang/global_config.py#L37
|
||||
FLASHINFER_WORKSPACE_BUFFER_SIZE = 394 * 1024 * 1024
|
||||
|
||||
|
||||
@@ -969,7 +969,7 @@ class TPUModelRunner(LoRAModelRunnerMixin):
|
||||
else:
|
||||
mm_embeds = []
|
||||
xm.mark_step()
|
||||
# Prepare inputs, the requests might be splitted into multiple
|
||||
# Prepare inputs, the requests might be split into multiple
|
||||
# executions, combine the result of each execution.
|
||||
start_index = 0
|
||||
combined_selected_tokens: list[torch.Tensor] = []
|
||||
|
||||
Reference in New Issue
Block a user