From bf27e34ebb5d2542b96647cffe42c3e8f9d83f16 Mon Sep 17 00:00:00 2001 From: Kyle Sayers Date: Thu, 23 Jul 2026 23:07:31 -0400 Subject: [PATCH] [CompressedTensors] DeepSeek4 CT Quantization Support (#41276) Signed-off-by: Kyle Sayers --- vllm/models/deepseek_v4/nvidia/ops/o_proj.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/models/deepseek_v4/nvidia/ops/o_proj.py b/vllm/models/deepseek_v4/nvidia/ops/o_proj.py index 18e3b10562..b064fbe9cd 100644 --- a/vllm/models/deepseek_v4/nvidia/ops/o_proj.py +++ b/vllm/models/deepseek_v4/nvidia/ops/o_proj.py @@ -60,10 +60,13 @@ def deep_gemm_fp8_o_proj( device=o.device, dtype=torch.bfloat16, ) + weight_scale = ( + wo_a.weight_scale if hasattr(wo_a, "weight_scale") else wo_a.weight_scale_inv + ) fp8_einsum( "bhr,hdr->bhd", (o_fp8, o_scale), - (wo_a.weight, wo_a.weight_scale_inv), + (wo_a.weight, weight_scale), z, recipe=einsum_recipe, )