vrr/kvcache-ai-ktransformers
2
0
Fork 0
mirror of https://github.com/kvcache-ai/ktransformers.git synced 2025-09-10 06:14:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

support npu

Browse source
This commit is contained in:
djw 2025-07-21 12:26:14 +00:00
parent dd0e41b3b8
commit 7d51a13c9b
34 changed files with 14004 additions and 5626 deletions
Download patch file Download diff file Expand all files Collapse all files

114
ktransformers/optimize/optimize_rules/npu/DeepSeek-V3-Chat-800IA2-npu.yaml Normal file
View file

@ -0,0 +1,114 @@
- match:
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3RotaryEmbedding
replace:
class: ktransformers.operators.RoPE.YarnRotaryEmbeddingV3
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
name: "^lm_head$" # regular expression
class: torch.nn.Linear # only match modules matching name and class simultaneously
replace:
class: ktransformers.operators.ascend.ascend_linear.KTransformersLinearW8A8A2 # optimized Kernel on quantized data types
kwargs:
generate_device: "npu"
prefill_device: "npu"
generate_op: "KLinearTorchW8A8A2"
prefill_op: "KLinearTorchW8A8A2"
- match:
name: "^model\\.layers\\.(?!.*self_attn\\.kv_b_proj).*$" # regular expression
class: torch.nn.Linear # only match modules matching name and class simultaneously
replace:
class: ktransformers.operators.ascend.ascend_linear.KTransformersLinearW8A8A2 # optimized Kernel on quantized data types
kwargs:
generate_device: "npu"
prefill_device: "npu"
generate_op: "KLinearTorchW8A8A2"
prefill_op: "KLinearTorchW8A8A2"
- match:
name: "^model\\.layers\\..*\\.mlp$"
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
replace:
class: ktransformers.operators.ascend.ascend_experts.KDeepseekV3MoEW8A8 # mlp module with custom forward function
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
name: "^model\\.layers\\.([0-2])\\.mlp$"
class: "ktransformers.models.modeling_deepseek_v3.DeepseekV3MLP"
replace:
class: "ktransformers.operators.ascend.ascend_mlp.KDeepseekV3MLPW8A8A2V1"
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
name: "^model\\.layers\\..*\\.mlp\\.shared_experts$"
class: "ktransformers.models.modeling_deepseek_v3.DeepseekV3MLP"
replace:
class: "ktransformers.operators.ascend.ascend_mlp.KDeepseekV3MLPW8A8A2V2"
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
class: ktransformers.models.modeling_deepseek_v3.MoEGate
replace:
class: ktransformers.operators.ascend.ascend_gate.KDeepseekV3GateA2
kwargs:
generate_device: "npu:0"
prefill_device: "npu:0"
- match:
name: "^model\\.layers\\..*\\.mlp\\.experts$"
replace:
class: ktransformers.operators.ascend.ascend_experts.KTransformersExpertsW8A8
kwargs:
prefill_device: "npu"
prefill_op: "KExpertsTorch"
generate_device: "cpu"
generate_op: "KExpertsCPUW8A8"
out_device: "npu"
recursive: False # don't recursively inject submodules of this module
- match:
name: "^model\\.layers\\..*\\.mlp\\.experts$"
class: ktransformers.operators.experts.KExpertsCPU
replace:
class: ktransformers.operators.ascend.ascend_experts.KExpertsCPUW8A8
- match:
name: "^model\\.layers\\..*\\.self_attn$"
replace:
class: ktransformers.operators.ascend.ascend_attention.KDeepseekV2AttentionW8A8A2 # optimized MLA implementation
kwargs:
generate_device: "npu"
prefill_device: "npu"
absorb_for_prefill: False # change this to True to enable long context(prefill may slower).
- match:
name: "^model$"
replace:
class: "ktransformers.operators.models.KDeepseekV2Model"
kwargs:
per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
- match:
name: "^model.embed_tokens"
replace:
class: "default"
kwargs:
generate_device: "cpu"
prefill_device: "cpu"
- match:
name: "^model..*norm"
replace:
class: ktransformers.operators.ascend.ascend_layernorm.KDeepseekV3RMSNormW8A8
kwargs:
generate_device: "npu"
prefill_device: "npu"

76
ktransformers/optimize/optimize_rules/npu/DeepSeek-V3-Chat-npu.yaml Normal file
View file

@ -0,0 +1,76 @@
- match:
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3RotaryEmbedding
replace:
class: ktransformers.operators.RoPE.YarnRotaryEmbeddingV3
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
name: "^lm_head$" # regular expression
class: torch.nn.Linear # only match modules matching name and class simultaneously
replace:
class: ktransformers.operators.linear.KTransformersLinear # optimized Kernel on quantized data types
kwargs:
generate_device: "npu"
prefill_device: "npu"
generate_op: "KLinearTorch"
prefill_op: "KLinearTorch"
- match:
name: "^model\\.layers\\.(?!.*self_attn\\.kv_b_proj).*$" # regular expression
class: torch.nn.Linear # only match modules matching name and class simultaneously
replace:
class: ktransformers.operators.linear.KTransformersLinear # optimized Kernel on quantized data types
kwargs:
generate_device: "npu"
prefill_device: "npu"
generate_op: "KLinearTorch"
prefill_op: "KLinearTorch"
- match:
name: "^model\\.layers\\..*\\.mlp$"
class: ktransformers.models.modeling_deepseek_v3.DeepseekV3MoE
replace:
class: ktransformers.operators.experts.KDeepseekV3MoE # mlp module with custom forward function
kwargs:
generate_device: "npu"
prefill_device: "npu"
- match:
class: ktransformers.models.modeling_deepseek_v3.MoEGate
replace:
class: ktransformers.operators.gate.KMoEGate
kwargs:
generate_device: "npu:0"
prefill_device: "npu:0"
- match:
name: "^model\\.layers\\..*\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersExperts # custom MoE Kernel with expert paralleism
kwargs:
prefill_device: "npu"
prefill_op: "KExpertsTorch"
generate_device: "cpu"
generate_op: "KExpertsCPU"
out_device: "npu"
recursive: False # don't recursively inject submodules of this module
- match:
name: "^model\\.layers\\..*\\.self_attn$"
replace:
class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
kwargs:
generate_device: "npu"
prefill_device: "npu"
absorb_for_prefill: False # change this to True to enable long context(prefill may slower).
- match:
name: "^model$"
replace:
class: "ktransformers.operators.models.KDeepseekV2Model"
kwargs:
per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
- match:
name: "^model.embed_tokens"
replace:
class: "default"
kwargs:
generate_device: "cpu"
prefill_device: "cpu"