[ROCm][CI] Minimize comment in RocmAttention q_scale check (#47731)
Signed-off-by: stefankoncarevic <stefan.koncarevic@amd.com>
stefankoncarevic committed
Jul 8, 2026 at 02:16 UTC
6e35c5e5af403b3d0e7ba3b56bd9733ca4480c31
1 file changed
+2
-7
vllm/v1/attention/backends/rocm_attn.py
+2
-7
@@ -421,13 +421,8 @@ class RocmAttentionImpl(AttentionImpl):
421
if is_quantized_kv_cache(self.kv_cache_dtype):
422
key_cache = key_cache.view(self.fp8_dtype)
423
value_cache = value_cache.view(self.fp8_dtype)
424
- # chunked_prefill_paged_decode runs attention with a full-precision
425
- # query (it does not quantize Q to fp8 and does not consume
426
- # q_scale), so q_scale only matters when the query itself is fp8.
427
- # For a non-fp8 query, q_scale is not applicable and is ignored
428
- # (mirrors TritonAttentionImpl). This avoids spuriously failing on
429
- # checkpoints that carry a non-1.0 q_scale while keeping the query
430
- # in full precision.
424
+ # q_scale only applies to an fp8 query; this path keeps the query
425
+ # in full precision, so a non-1.0 q_scale is not applicable here.
426
if query.dtype == self.fp8_dtype and layer._q_scale_float != 1.0:
427
raise NotImplementedError(
428
"A non 1.0 q_scale with an fp8 query is not currently "