EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[QUANT] Add GPTQModel Dynamic Quantization + lm_head Quantization (#3790)

Browse Source 
Signed-off-by: ZX-ModelCloud <zx@modelcloud.ai>
Co-authored-by: ZX-ModelCloud <zx@modelcloud.ai>
This commit is contained in:
Qubitium-ModelCloud
2025-03-05 17:11:00 +08:00
committed by GitHub
parent 583d6af71b
commit 56a724eba3
56 changed files with 1988 additions and 282 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
python/sglang/srt/models/llava.py
View File

@@ -42,6 +42,7 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.models.llama import LlamaForCausalLM
from sglang.srt.models.mistral import MistralForCausalLM
from sglang.srt.models.qwen2 import Qwen2ForCausalLM
from sglang.srt.utils import add_prefix
class LlavaBaseForCausalLM(nn.Module):
@@ -475,6 +476,7 @@ class LlavaLlamaForCausalLM(LlavaBaseForCausalLM):
self,
config: LlavaConfig,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
@@ -484,7 +486,11 @@ class LlavaLlamaForCausalLM(LlavaBaseForCausalLM):
self.config.text_config.hidden_size = config.hidden_size
self.multi_modal_projector = LlavaMultiModalProjector(config)
self.language_model = LlamaForCausalLM(config, quant_config=quant_config)
self.language_model = LlamaForCausalLM(
config,
quant_config=quant_config,
prefix=add_prefix("language_model", prefix),
)
if "unpad" in getattr(config, "mm_patch_merge_type", ""):
self.language_model.model.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size, dtype=torch.float16)
@@ -496,6 +502,7 @@ class LlavaQwenForCausalLM(LlavaBaseForCausalLM):
self,
config: LlavaConfig,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
@@ -516,7 +523,11 @@ class LlavaQwenForCausalLM(LlavaBaseForCausalLM):
self.config.image_token_index = 151646
self.multi_modal_projector = LlavaMultiModalProjector(config)
self.language_model = Qwen2ForCausalLM(config, quant_config=quant_config)
self.language_model = Qwen2ForCausalLM(
config,
quant_config=quant_config,
prefix=add_prefix("language_model", prefix),
)
if "unpad" in getattr(config, "mm_patch_merge_type", ""):
self.language_model.model.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size, dtype=torch.float16)
@@ -528,6 +539,7 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
self,
config: LlavaConfig,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
) -> None:
super().__init__()
@@ -548,7 +560,11 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
self.config.image_token_index = 32000
self.multi_modal_projector = LlavaMultiModalProjector(config)
self.language_model = MistralForCausalLM(config, quant_config=quant_config)
self.language_model = MistralForCausalLM(
config,
quant_config=quant_config,
prefix=add_prefix("language_model", prefix),
)
if "unpad" in getattr(config, "mm_patch_merge_type", ""):
self.language_model.model.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size, dtype=torch.float16)