[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:
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user