[CORE] Quantized lm-head Framework (#4442)
Co-authored-by: Robert Shaw <rshaw@neuralmagic.com> Co-authored-by: ZX <zx@lbx.dev>
This commit is contained in:
co-authored by
Robert Shaw
ZX
parent
7c008c51a9
commit
ee93f4f92a
@@ -331,6 +331,7 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA):
|
||||
# We need bigger padding if using lora for kernel
|
||||
# compatibility
|
||||
if not lora_config else lora_config.lora_vocab_padding_size,
|
||||
quant_config=quant_config,
|
||||
)
|
||||
self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
|
||||
config.vocab_size)
|
||||
@@ -350,7 +351,7 @@ class MixtralForCausalLM(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
|
||||
|
||||
|
||||
Reference in New Issue
Block a user