423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-07-28 19:56:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Jiaqi
2026-07-24 15:23:21 +08:00
committed by GitHub
parent 19e9fe3ced
commit 3f77101580
50 changed files with 843 additions and 1007 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests_v1/plugins/model_plugins/test_kernel_plugin.py
View File

@@ -30,13 +30,13 @@ def _apply_kernel(rank) -> None:
if k.startswith("llamafactory.v1.plugins.model_plugins.kernels"):
del sys.modules[k]
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_default_kernels
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_kernels
model = AutoModelForCausalLM.from_pretrained("llamafactory/tiny-random-qwen3")
original_rmsnorm_forward = model.model.layers[0].input_layernorm.forward
original_swiglu_forward = model.model.layers[0].mlp.forward
model = apply_default_kernels(model=model, include_kernels="npu_fused_rmsnorm")
model = apply_kernels(model=model, config={"name": "npu_fused_rmsnorm"})
assert model.model.layers[0].input_layernorm.forward.__func__ is not original_rmsnorm_forward.__func__
assert model.model.layers[0].mlp.forward.__func__ is original_swiglu_forward.__func__
@@ -53,13 +53,13 @@ def _apply_all_kernels(rank) -> None:
if k.startswith("llamafactory.v1.plugins.model_plugins.kernels"):
del sys.modules[k]
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_default_kernels
from llamafactory.v1.plugins.model_plugins.kernels.interface import apply_kernels
model = AutoModelForCausalLM.from_pretrained("llamafactory/tiny-random-qwen3")
original_rmsnorm_forward = model.model.layers[0].input_layernorm.forward
original_swiglu_forward = model.model.layers[0].mlp.forward
model = apply_default_kernels(model=model, include_kernels=True)
model = apply_kernels(model=model, config={"name": "auto"})
assert model.model.layers[0].input_layernorm.forward.__func__ is not original_rmsnorm_forward.__func__
assert model.model.layers[0].mlp.forward.__func__ is not original_swiglu_forward.__func__

8
tests_v1/plugins/model_plugins/test_ulysses_cp.py
View File

@@ -18,6 +18,7 @@ import torch.multiprocessing as mp
from llamafactory.v1.accelerator.interface import DistributedInterface
from llamafactory.v1.config.model_args import ModelArguments
from llamafactory.v1.config.training_args import TrainingArguments
from llamafactory.v1.core.model_engine import ModelEngine
from llamafactory.v1.plugins.model_plugins.parallelization.sequence_parallel import (
SequenceParallelModelPlugin,
@@ -33,15 +34,14 @@ def _test_sequence_parallel_loss(
with dist_env(local_rank, world_size, master_port):
model_args = ModelArguments(model="llamafactory/tiny-random-qwen3")
# Initialize distributed interface with config
dist_config = {"cp_mode": "ulysses", "cp_size": cp_size, "dp_size": dp_size}
DistributedInterface(dist_config)
training_args = TrainingArguments(cp_mode="ulysses", cp_size=cp_size, dp_size=dp_size)
DistributedInterface(training_args)
# Now create model engine
model_engine = ModelEngine(model_args=model_args)
# Apply sequence parallel plugin
SequenceParallelModelPlugin(dist_config.get("cp_mode", "ulysses"))(model_engine.model, dist_config)
SequenceParallelModelPlugin(training_args.cp_mode)(model_engine.model, training_args.cp_size)
input_ids = torch.arange(1, batch_size * 5 + 1, dtype=torch.long).view(batch_size, 5)
model_inputs = {