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