diff --git a/vllm/models/deepseek_v4/xpu/xpu_qnorm_rope_kv_fp8_insert.py b/vllm/models/deepseek_v4/xpu/xpu_qnorm_rope_kv_fp8_insert.py index a6cd4fbd33..4d4ab80a62 100644 --- a/vllm/models/deepseek_v4/xpu/xpu_qnorm_rope_kv_fp8_insert.py +++ b/vllm/models/deepseek_v4/xpu/xpu_qnorm_rope_kv_fp8_insert.py @@ -88,10 +88,13 @@ def _xpu_qnorm_rope_kernel( kv_base = kv_ptr + token_idx * HEAD_DIM kv_out_base = kv_out_ptr + token_idx * HEAD_DIM - # Copy full KV unchanged first + # Copy only the NoPE portion; the RoPE region is written exclusively + # by the rotated stores below. Masking avoids a write-after-write + # hazard on the overlapping RoPE addresses (matches the Q branch). offs = tl.arange(0, HEAD_DIM) kv_full = tl.load(kv_base + offs) - tl.store(kv_out_base + offs, kv_full) + nope_mask = offs < NOPE_DIM + tl.store(kv_out_base + offs, kv_full, mask=nope_mask) # GPT-J interleaved RoPE on the last ROPE_DIM dimensions even_offs = NOPE_DIM + rope_pair_idx * 2