[Perf] Avoid redundant conversions in FP8 dummy initialization (#56688)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai> Co-authored-by: OpenAI Codex <noreply@openai.com>
Woosuk Kwon committed
Sep 13, 2026 at 12:06 UTC
71888f507ad5df5af2fae2ab39a6c1b0cc465eb1
1 file changed
+3
-4
vllm/model_executor/model_loader/weight_utils.py
+3
-4
@@ -1380,10 +1380,9 @@ def initialize_single_dummy_weight(
1380
generator.manual_seed(seed)
1381
if torch.finfo(param.data.dtype).bits < 16:
1382
# uniform_ doesn't support < 16-bit datatypes (FP8)
1383
- dtype = param.data.dtype
1384
- tmp_param = param.data.to(torch.float16)
1385
- tmp_param = tmp_param.uniform_(low, high, generator=generator).to(dtype)
1386
- param.data.copy_(tmp_param)
1383
+ tmp_param = torch.empty_like(param, dtype=torch.float16)
1384
+ tmp_param.uniform_(low, high, generator=generator)
1385
+ param.copy_(tmp_param)
1386
else:
1387
param.uniform_(low, high, generator=generator)
1388