[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>
This commit is contained in:
co-authored by
OpenAI Codex
parent
b6e2aa748b
commit
71888f507a
@@ -1380,10 +1380,9 @@ def initialize_single_dummy_weight(
|
||||
generator.manual_seed(seed)
|
||||
if torch.finfo(param.data.dtype).bits < 16:
|
||||
# uniform_ doesn't support < 16-bit datatypes (FP8)
|
||||
dtype = param.data.dtype
|
||||
tmp_param = param.data.to(torch.float16)
|
||||
tmp_param = tmp_param.uniform_(low, high, generator=generator).to(dtype)
|
||||
param.data.copy_(tmp_param)
|
||||
tmp_param = torch.empty_like(param, dtype=torch.float16)
|
||||
tmp_param.uniform_(low, high, generator=generator)
|
||||
param.copy_(tmp_param)
|
||||
else:
|
||||
param.uniform_(low, high, generator=generator)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user