Skip to content

GitLab

Commit 7f1d5aff authored by zhuwenwen's avatar zhuwenwen
Browse files

Revert "update kv_b_proj_weight" 

This reverts commit 3b121add.
parent 3b121add
Hide whitespace changes
Inline Side-by-side
Showing with 1 addition and 4 deletions
vllm/v1/attention/backends/mla/common.py
View file @ 7f1d5aff
...@@ -1102,10 +1102,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]): ...@@ -1102,10 +1102,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
# we currently do not have quantized bmm's which are needed for # we currently do not have quantized bmm's which are needed for
# `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform
# the bmm's in 16-bit, the extra memory overhead of this is fairly low # the bmm's in 16-bit, the extra memory overhead of this is fairly low
if self.use_llama_nn and isinstance(self.kv_b_proj.quant_method, UnquantizedLinearMethod): kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj)
else:
kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
assert kv_b_proj_weight.shape == ( assert kv_b_proj_weight.shape == (
self.kv_lora_rank, self.kv_lora_rank,
self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), ( self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), (
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment