[v1] refactor registry plugin structure and params (#10641)

2026-07-28 11:46:09 +08:00 · 2026-07-24 15:23:21 +08:00
parent 19e9fe3ced
commit 3f77101580
50 changed files with 843 additions and 1007 deletions
--- a/examples/v1/train_batching_strategy/train_full_fsdp2_batching_normal.yaml
+++ b/examples/v1/train_batching_strategy/train_full_fsdp2_batching_normal.yaml
@@ -1,12 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-template: qwen3_nothink
-
-
 kernel_config:
  name: auto
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_batching_strategy/train_full_fsdp2_dynamic_batching.yaml
+++ b/examples/v1/train_batching_strategy/train_full_fsdp2_dynamic_batching.yaml
@@ -1,11 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-template: qwen3_nothink
-
 kernel_config:
  name: auto
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_batching_strategy/train_full_fsdp2_dynamic_padding_free.yaml
+++ b/examples/v1/train_batching_strategy/train_full_fsdp2_dynamic_padding_free.yaml
@@ -1,11 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-template: qwen3_nothink
-
 kernel_config:
  name: auto
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_batching_strategy/train_full_fsdp2_padding_free.yaml
+++ b/examples/v1/train_batching_strategy/train_full_fsdp2_padding_free.yaml
@@ -1,11 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-template: qwen3_nothink
-
 kernel_config:
  name: auto
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_freeze/train_freeze_sft.yaml
+++ b/examples/v1/train_freeze/train_freeze_sft.yaml
@@ -1,7 +1,6 @@
 model: Qwen/Qwen3-4B
 model_class: llm

-
 # Freeze Configuration
 peft_config:
  name: freeze
@@ -12,7 +11,6 @@ peft_config:
 # Kernel Config
 kernel_config:
  name: auto
-  include_kernels: auto

 # FSDP Config
 dist_config:
@@ -25,7 +23,6 @@ train_dataset: data/v1_sft_demo.yaml
 ### training
 output_dir: ./outputs/test_freeze
 micro_batch_size: 1
-global_batch_size: 4
 cutoff_len: 2048
 learning_rate: 2.0e-5
 max_steps: 10
--- a/examples/v1/train_full/train_full_deepspeed.yaml
+++ b/examples/v1/train_full/train_full_deepspeed.yaml
@@ -1,10 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-
 kernel_config:
    name: auto
-    include_kernels: auto

 dist_config:
    name: deepspeed
--- a/examples/v1/train_full/train_full_fsdp2.yaml
+++ b/examples/v1/train_full/train_full_fsdp2.yaml
@@ -1,10 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-
 kernel_config:
  name: auto
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_full/train_full_liger_kernel.yaml
+++ b/examples/v1/train_full/train_full_liger_kernel.yaml
@@ -1,11 +1,8 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-template: qwen3_nothink
-
 kernel_config:
  name: liger_kernel
-  include_kernels: auto # choice: null/true/false/auto/kernel_id1,kernel_id2,kernel_id3, default is null

 quant_config: null

--- a/examples/v1/train_full/train_full_ulysses_cp.yaml
+++ b/examples/v1/train_full/train_full_ulysses_cp.yaml
@@ -8,8 +8,9 @@ flash_attn: flash_attention_2
 dist_config:
  name: fsdp2
  dcp_path: null
-  cp_mode: ulysses
-  cp_size: 2
+
+cp_mode: ulysses
+cp_size: 2

 ### data
 train_dataset: data/v1_sft_demo.yaml
--- a/examples/v1/train_lora/train_lora_dpo.yaml
+++ b/examples/v1/train_lora/train_lora_dpo.yaml
@@ -1,8 +1,6 @@
 model: Qwen/Qwen3-4B
 model_class: llm

-template: qwen3_nothink
-
 # PEFT Configuration
 peft_config:
  name: lora
@@ -14,7 +12,6 @@ peft_config:
 # Kernel Config
 kernel_config:
  name: auto
-  include_kernels: auto

 # FSDP Config
 dist_config:
--- a/examples/v1/train_lora/train_lora_sft.yaml
+++ b/examples/v1/train_lora/train_lora_sft.yaml
@@ -1,7 +1,6 @@
 model: Qwen/Qwen3-4B
 model_class: llm

-
 # PEFT Configuration
 peft_config:
  name: lora
@@ -13,7 +12,6 @@ peft_config:
 # Kernel Config
 kernel_config:
  name: auto
-  include_kernels: auto

 # FSDP Config
 dist_config:
--- a/examples/v1/train_lora/train_lora_sft_rank0.yaml
+++ b/examples/v1/train_lora/train_lora_sft_rank0.yaml
@@ -1,7 +1,6 @@
 model: Qwen/Qwen3-4B
 model_class: llm

-
 # PEFT Configuration
 peft_config:
  name: lora
@@ -13,7 +12,6 @@ peft_config:
 # Kernel Config
 kernel_config:
  name: auto
-  include_kernels: auto

 # FSDP Config
 dist_config:
--- a/examples/v1/train_qlora/quantization.yaml
+++ b/examples/v1/train_qlora/quantization.yaml
@@ -1,7 +1,6 @@
 model: Qwen/Qwen3-0.6B
 model_class: llm

-
 # PEFT Configuration
 peft_config:
  name: lora
@@ -13,7 +12,6 @@ peft_config:
 # Kernel Config
 kernel_config:
  name: auto
-  include_kernels: auto

 # FSDP Config
 dist_config: