biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CORE] Quantized lm-head Framework (#4442)

Browse Source 
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com>
Co-authored-by: ZX <zx@lbx.dev>
This commit is contained in:
Qubitium-ModelCloud
2024-07-03 06:25:17 +08:00
committed by GitHub
parent 7c008c51a9
commit ee93f4f92a
48 changed files with 268 additions and 121 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/models/gpt_bigcode.py
View File

@@ -259,7 +259,7 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA):
self.quant_config = quant_config
self.transformer = GPTBigCodeModel(config, cache_config, quant_config,
lora_config)
self.lm_head_weight = self.transformer.wte.weight
self.lm_head = self.transformer.wte
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
@@ -281,7 +281,7 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA):
def compute_logits(self, hidden_states: torch.Tensor,
sampling_metadata: SamplingMetadata) -> torch.Tensor:
logits = self.logits_processor(self.lm_head_weight, hidden_states,
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
return logits