vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2025-09-09 13:55:27 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

support qwen3, dont speak human language

Browse source
This commit is contained in:
djw 2025-04-28 08:44:47 +00:00
parent f3d842a0ca
commit 3f9bbf1181
30 changed files with 3696 additions and 290 deletions
Download patch file Download diff file Expand all files Collapse all files

18
ktransformers/operators/mlp.py
View file

@ -4,8 +4,7 @@ from ktransformers.util.custom_gguf import GGUFLoader
from transformers import PretrainedConfig
import torch.nn as nn
from ktransformers.models.modeling_deepseek_v3 import DeepseekV3MLP
from ktransformers.models.modeling_qwen2_moe import Qwen2MoeMLP
class kDeepseekV3MLP(DeepseekV3MLP, BaseInjectedModule):
def __init__(self,
key: str,
@ -18,6 +17,21 @@ class kDeepseekV3MLP(DeepseekV3MLP, BaseInjectedModule):
BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
self.orig_module.__init__(orig_module.config,
orig_module.hidden_size, orig_module.intermediate_size)
def forward(self, x, bsz_tensor):
down_proj = self.down_proj(self.act_fn(self.gate_proj(x, bsz_tensor)) * self.up_proj(x, bsz_tensor), bsz_tensor)
return down_proj
class KQwen2MoeMLP(Qwen2MoeMLP, BaseInjectedModule):
def __init__(self,
key: str,
gguf_loader : GGUFLoader,
config: PretrainedConfig,
orig_module: nn.Module,
prefill_device: str = "cuda",
generate_device: str = "cuda",
**kwargs):
BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, **kwargs)
self.orig_module.__init__(orig_module.config,
orig_module.intermediate_size)
def forward(self, x, bsz_tensor):
down_proj = self.down_proj(self.act_fn(self.gate_proj(x, bsz_tensor)) * self.up_proj(x, bsz_tensor), bsz_tensor)
return down_proj