Skip to content

[Bugfix] Fix MLA attention crash with AWQ/GPTQ quantized models #34695

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
MatthewBonanni merged 3 commits into from
Mar 13, 2026
Merged

[Bugfix] Fix MLA attention crash with AWQ/GPTQ quantized models #34695

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
73c6bff
[Bugfix] Fix MLA attention crash with AWQ/GPTQ quantized models (#34561)
haosdent Feb 17, 2026
0f6b38a
Merge branch 'main' into fix-34561
MatthewBonanni Mar 6, 2026
88b81aa
Merge branch 'main' into fix-34561
pavanimajety Mar 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 10 additions & 5 deletions vllm/model_executor/layers/attention/mla_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -442,6 +442,7 @@ def __init__(
# If kv_b_proj_weight is unquantized, quantize it to mxfp4 if supported
self.is_aiter_triton_fp4_bmm_enabled = (
rocm_aiter_ops.is_fp4bmm_enabled()
and hasattr(self.kv_b_proj, "weight")
and self.kv_b_proj.weight.dtype == torch.bfloat16
)

Expand Down Expand Up @@ -2492,11 +2493,15 @@ def _compute_prefill_context(
kv_c_normed = workspace[:toks][..., : self.kv_lora_rank]
# When FP8 weights are used without FP8 prefill, kv_b_proj expects
# model dtype input and will quantize internally.
if (
use_fp8_prefill
or self.kv_b_proj.weight.dtype != current_platform.fp8_dtype()
):
kv_c_normed = kv_c_normed.to(self.kv_b_proj.weight.dtype)
# For quantized layers (AWQ/GPTQ) that lack a .weight attribute,
# use params_dtype which is the expected input dtype.
_kv_b_proj_w_dtype = (
self.kv_b_proj.weight.dtype
if hasattr(self.kv_b_proj, "weight")
else self.kv_b_proj.params_dtype
)
if use_fp8_prefill or _kv_b_proj_w_dtype != current_platform.fp8_dtype():
kv_c_normed = kv_c_normed.to(_kv_b_proj_w_dtype)

k_pe = workspace[:toks][..., self.kv_lora_rank :].unsqueeze(1)
kv_nope = self.kv_b_proj(kv_c_normed)[0].view(
Expand Down
Loading