Mirrors/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-12-23 21:18:00 +01:00
Code Issues Packages Projects Releases Wiki Activity

Fix Multi-GPU not working on exllama_hf (#2803)

Browse Source
This commit is contained in:
Panchovix 2023-06-22 15:05:25 -04:00 committed by GitHub
parent d94ea31d54
commit b4a38c24b7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 8 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
modules/exllama_hf.py
View File

@ -38,7 +38,6 @@ class ExllamaHF(PreTrainedModel):
@property
def device(self) -> torch.device:
# TODO: May cause problem on multi-gpu inference?
return torch.device(0)
def __call__(self, *args, **kwargs):
@ -50,7 +49,7 @@ class ExllamaHF(PreTrainedModel):
if cache is None:
cache = ExLlamaCache(self.ex_model)
self.ex_model.forward(torch.tensor([seq[:-1]], dtype=torch.long), cache, preprocess_only=True)
logits = self.ex_model.forward(torch.tensor([seq[-1:]], dtype=torch.long), cache).to(self.device)
logits = self.ex_model.forward(torch.tensor([seq[-1:]], dtype=torch.long), cache).to(kwargs['input_ids'].device)
return CausalLMOutputWithPast(logits=logits, past_key_values=cache if use_cache else None)
@classmethod
@ -72,11 +71,14 @@ class ExllamaHF(PreTrainedModel):
assert weight_path is not None, f'could not find weight in "{pretrained_model_name_or_path}"'
config.model_path = str(weight_path)
if shared.args.gpu_split:
config.set_auto_map(shared.args.gpu_split)
config.gpu_peer_fix = True
# This slowes down a bit but align better with autogptq generation.
# TODO: Should give user choice to tune the exllama config
config.act_order = True
config.fused_attn = False
config.fused_mlp_thd = 0
# config.fused_attn = False
# config.fused_mlp_thd = 0
return ExllamaHF(config)
return ExllamaHF(config)