[CompressedTensors] DeepSeek4 CT Quantization Support (#41276)
Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
Kyle Sayers committed
Jul 23, 2026 at 23:07 UTC
bf27e34ebb5d2542b96647cffe42c3e8f9d83f16
1 file changed
+4
-1
vllm/models/deepseek_v4/nvidia/ops/o_proj.py
+4
-1
@@ -60,10 +60,13 @@ def deep_gemm_fp8_o_proj(
60
device=o.device,
61
dtype=torch.bfloat16,
62
)
63
+ weight_scale = (
64
+ wo_a.weight_scale if hasattr(wo_a, "weight_scale") else wo_a.weight_scale_inv
65
+ )
66
fp8_einsum(
67
"bhr,hdr->bhd",
68
(o_fp8, o_scale),
66
- (wo_a.weight, wo_a.weight_scale_inv),
69
+ (wo_a.weight, weight_scale),
70
z,
71
recipe=einsum_recipe,
72
)