[Refactor][Quantization] Make FP8 and MLA weight transforms reusable pure functions (#57732)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
@@ -404,6 +404,24 @@ def _get_kv_b_proj_input_dtype(
|
||||
return weight_dtype
|
||||
|
||||
|
||||
def split_kv_b_proj(
|
||||
kv_b_proj: nn.Module,
|
||||
out_dtype: torch.dtype,
|
||||
kv_lora_rank: int,
|
||||
num_heads: int,
|
||||
qk_nope_head_dim: int,
|
||||
v_head_dim: int,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Dequantize ``kv_b_proj`` and return ``W_UK [L,N,P]``, ``W_UV [L,N,V]``."""
|
||||
weight = get_and_maybe_dequant_weights(kv_b_proj, out_dtype=out_dtype).T
|
||||
assert weight.shape == (
|
||||
kv_lora_rank,
|
||||
num_heads * (qk_nope_head_dim + v_head_dim),
|
||||
), f"kv_b_proj weight {tuple(weight.shape)} vs {kv_lora_rank=} {num_heads=}"
|
||||
weight = weight.view(kv_lora_rank, num_heads, qk_nope_head_dim + v_head_dim)
|
||||
return weight.split([qk_nope_head_dim, v_head_dim], dim=-1)
|
||||
|
||||
|
||||
class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
"""Multi-Head Latent Attention layer.
|
||||
|
||||
@@ -1221,13 +1239,6 @@ class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
)
|
||||
return
|
||||
|
||||
# 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
|
||||
# the bmm's in 16-bit, the extra memory overhead of this is fairly low
|
||||
kv_b_proj_weight = get_and_maybe_dequant_weights(
|
||||
self.kv_b_proj, out_dtype=act_dtype
|
||||
).T
|
||||
|
||||
if self.dcp_q_replicate:
|
||||
# qrep wired here: validate unsupported decode backends once.
|
||||
assert self.q_pad_num_heads in (None, self.num_heads), (
|
||||
@@ -1243,24 +1254,13 @@ class MLAAttention(nn.Module, AttentionLayerBase):
|
||||
"FP4/FP8 MLA BMM paths."
|
||||
)
|
||||
|
||||
assert kv_b_proj_weight.shape == (
|
||||
self.kv_lora_rank,
|
||||
self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
|
||||
), (
|
||||
f"{kv_b_proj_weight.shape=}, "
|
||||
f"{self.kv_lora_rank=}, "
|
||||
f"{self.num_heads=}, "
|
||||
f"{self.qk_nope_head_dim=}, "
|
||||
f"{self.v_head_dim=}"
|
||||
)
|
||||
kv_b_proj_weight = kv_b_proj_weight.view(
|
||||
W_UK, W_UV = split_kv_b_proj(
|
||||
self.kv_b_proj,
|
||||
act_dtype,
|
||||
self.kv_lora_rank,
|
||||
self.num_heads,
|
||||
self.qk_nope_head_dim + self.v_head_dim,
|
||||
)
|
||||
|
||||
W_UK, W_UV = kv_b_proj_weight.split(
|
||||
[self.qk_nope_head_dim, self.v_head_dim], dim=-1
|
||||
self.qk_nope_head_dim,
|
||||
self.v_head_dim,
|
||||
)
|
||||
|
||||
# If kv_b_proj_weight is unquantized, quantize it to mxfp4 if supported
|
||||
|
||||
Reference in New Issue
Block a user