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

5
vllm/model_executor/models/llava_next.py
View File

@@ -186,7 +186,8 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsVision):
self.lm_head = ParallelLMHead(
self.unpadded_vocab_size,
config.text_config.hidden_size,
org_num_embeddings=self.language_model.org_vocab_size)
org_num_embeddings=self.language_model.org_vocab_size,
quant_config=quant_config)
logit_scale = getattr(config, "logit_scale", 1.0)
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
config.vocab_size, logit_scale)
@@ -438,7 +439,7 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsVision):
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