diff --git a/vllm/models/qwen4_exp/nvidia/ops/qsa.py b/vllm/models/qwen4_exp/nvidia/ops/qsa.py index c7fe601449..9230be24a8 100644 --- a/vllm/models/qwen4_exp/nvidia/ops/qsa.py +++ b/vllm/models/qwen4_exp/nvidia/ops/qsa.py @@ -28,10 +28,30 @@ def _maybe_dequant(tile): # e4m3 bytes below sm_89 (no Triton fp8e4nv pointer type): sign to bit # 15, exp+mantissa to bits 13..7 gives the fp16 of x * 2^-8. Exact for # every finite code; the two NaN codes decode to +-480 after the host - # scale instead of NaN. - code = tile.to(tl.uint16) - bits = ((code & 0x80) << 8) | ((code & 0x7F) << 7) - return bits.to(tl.float16, bitcast=True) + # scale instead of NaN. Decodes four bytes per 32-bit register: the + # rotate moves each byte's bits 6..1 to 5..0 and its bit 0 to bit 7 of + # the byte below, lop3 restores the sign bits, prmt pairs high and low + # bytes, and the mask clears the sign copy at bit 14 and bits 6..0. + # pack=4 puts element k in byte k of $2 and half k of ($0, $1), the + # same lane mapping as Triton's own sm<89 fp8 -> fp16 asm. + return tl.inline_asm_elementwise( + asm=""" + { + .reg .b32 rot, high; + shf.r.wrap.b32 rot, $2, $2, 1; + lop3.b32 high, $2, rot, 0x80808080, 0xE4; + prmt.b32 $0, rot, high, 0x5043; + prmt.b32 $1, rot, high, 0x7261; + and.b32 $0, $0, 0xBF80BF80; + and.b32 $1, $1, 0xBF80BF80; + } + """, + constraints="=r,=r,r", + args=[tile], + dtype=tl.float16, + is_pure=True, + pack=4, + ) elif tile.dtype == tl.float8e4nv: return tile.to(tl.float16) else: