vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2026-04-28 20:00:06 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

update rope calculation; update modeling.py; update gate for moe

Browse source
This commit is contained in:
Azure 2025-02-01 07:32:21 +00:00
parent 5a50b34627
commit f873558a89
11 changed files with 402 additions and 412 deletions
Download patch file Download diff file Expand all files Collapse all files

8
ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-multi-gpu.yaml
View file

@ -47,7 +47,7 @@
- match:
name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp$"
class: ktransformers.models.modeling_deepseekv3.DeepseekV3MoE
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
replace:
class: ktransformers.operators.experts.KDeepseekV3MoE # mlp module with custom forward function
kwargs:
@ -55,7 +55,7 @@
prefill_device: "cuda:0"
- match:
name: "^model\\.layers\\.([3456][0-9])\\.mlp$"
class: ktransformers.models.modeling_deepseekv3.DeepseekV3MoE
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
replace:
class: ktransformers.operators.experts.KDeepseekV3MoE # mlp module with custom forward function
kwargs:
@ -64,7 +64,7 @@
- match:
name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.mlp\\.gate$"
class: ktransformers.models.modeling_deepseekv3.MoEGate
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3TopkRouter
replace:
class: ktransformers.operators.gate.KMoEGate
kwargs:
@ -72,7 +72,7 @@
prefill_device: "cuda:0"
- match:
name: "^model\\.layers\\.([3456][0-9])\\.mlp\\.gate$"
class: ktransformers.models.modeling_deepseekv3.MoEGate
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3TopkRouter
replace:
class: ktransformers.operators.gate.KMoEGate # mlp module with custom forward function
kwargs: