[ci] update ci (#6254)

* fix for async io

* test for upgrading transformers

* add ci machine

* fix

* fix

* fix

* fix

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update test_fp16_torch.py

* Update build_on_pr.yml

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* fix

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* fiux

* fix

* fix

* fix

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

.github/workflows/build_on_pr.yml

@@ -87,10 +87,10 @@ jobs:
     name: Build and Test Colossal-AI
     needs: detect
     if: needs.detect.outputs.anyLibraryFileChanged == 'true'
-    runs-on: [self-hosted, gpu]
+    runs-on: ubuntu-latest
     container:
-      image: hpcaitech/pytorch-cuda:2.2.2-12.1.0
-      options: --gpus all --rm -v /dev/shm -v /data/scratch:/data/scratch
+      image: image-cloud.luchentech.com/hpcaitech/pytorch-cuda:2.2.2-12.1.0
+      options: --gpus all --shm-size=2g --rm -v /dev/shm -v /data/scratch:/data/scratch
     timeout-minutes: 90
     defaults:
       run:

README.md

@@ -38,7 +38,7 @@ Limited Academic Bonuses:
 
 <div align="center">
    <a href="https://hpc-ai.com/?utm_source=github&utm_medium=social&utm_campaign=promotion-colossalai">
-   <img src="https://github.com/hpcaitech/public_assets/blob/main/colossalai/img/2.gif" width="850" />
+   <img src="https://github.com/hpcaitech/public_assets/blob/main/colossalai/img/2-2.gif" width="850" />
    </a>
 </div>
 

applications/ColossalChat/examples/README.md

@@ -892,6 +892,63 @@ The dialogues can by multiple turns and it can contain system prompt. For more d
 
 We use bf16 weights for finetuning. If you downloaded fp8 DeepSeek V3/R1 weights, you can use the [script](https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/fp8_cast_bf16.py) to convert the weights to bf16 via GPU. For Ascend NPU, you can use this [script](https://gitee.com/ascend/ModelZoo-PyTorch/blob/master/MindIE/LLM/DeepSeek/DeepSeek-V2/NPU_inference/fp8_cast_bf16.py).
 
+We have also added details on how to load and reason with lora models.
+```python
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+)
+from peft import (
+    PeftModel
+)
+import torch
+
+# Set model path
+model_name = "Qwen/Qwen2.5-3B"
+lora_adapter = "Qwen2.5-3B_lora" # Your lora model Path
+merged_model_path = "Qwen2.5-3B_merged"
+
+######
+# How to Load lora Model
+######
+# 1.Load base model
+base_model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True
+)
+
+# 2.Load lora model
+peft_model = PeftModel.from_pretrained(
+    base_model,
+    lora_adapter,
+    torch_dtype=torch.bfloat16
+)
+
+# 3.Merge lora model
+merged_model = peft_model.merge_and_unload()
+
+# 4.Load tokenizer
+tokenizer = AutoTokenizer.from_pretrained(
+    model_name,
+    trust_remote_code=True,
+    pad_token="<|endoftext|>"
+)
+
+# 5.Save merged lora model
+merged_model.save_pretrained(
+    merged_model_path,
+    safe_serialization=True
+)
+tokenizer.save_pretrained(merged_model_path)
+
+# 6.Run Inference
+test_input = tokenizer("Instruction: Finding prime numbers up to 100\nAnswer:", return_tensors="pt").to("cuda")
+output = merged_model.generate(**test_input, max_new_tokens=100)
+print(tokenizer.decode(output[0], skip_special_tokens=True))
+```
+
 #### Usage
 
 After preparing the dataset and model weights, you can run the script with the following command:

tests/test_fp8/test_all_to_all_single.py

@@ -6,9 +6,10 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import all_to_all_single_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(4,), (1, 8, 16), (4, 8, 16)])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
 @parameterize("async_op", [True, False])
@@ -24,6 +25,7 @@ def check_all2all(shape, dtype, async_op):
     assert_close(output, output_fp8, rtol=0.1, atol=0.1)
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(8, 8, 16)])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
 @parameterize("async_op", [True, False])

tests/test_fp8/test_fp8_all_to_all.py

@@ -6,9 +6,10 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import _all_to_all_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(16, 8, 4)])
 @parameterize("scatter_dim", [0, 1, 2])
 @parameterize("dtype", [torch.bfloat16, torch.float16])

tests/test_fp8/test_fp8_all_to_all_single.py

@@ -6,11 +6,12 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import all_to_all_single_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 dist.all_to_all_single
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(4), (8, 7), (4, 8, 16)])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
 @parameterize("fp8_format", ["e4m3", "e5m2"])

tests/test_fp8/test_fp8_allgather.py

@@ -6,9 +6,10 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import _all_gather_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 
+@clear_cache_before_run()
 @parameterize(
     "shape",
     [(3, 7, 16)],

tests/test_fp8/test_fp8_allreduce.py

@@ -5,7 +5,7 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import all_reduce_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 
 @parameterize(
@@ -20,6 +20,7 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
         (8,),
     ],
 )
+@clear_cache_before_run()
 @parameterize("dtype", [torch.float16, torch.bfloat16])
 @parameterize("fp8_format", ["e4m3", "e5m2"])
 @parameterize("async_op", [True, False])

tests/test_fp8/test_fp8_cast.py

@@ -3,9 +3,10 @@ from torch.testing import assert_close
 
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import cast_from_fp8, cast_from_fp8_pipeline, cast_to_fp8, cast_to_fp8_pipeline
-from colossalai.testing import parameterize
+from colossalai.testing import clear_cache_before_run, parameterize
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(100, 10), (10, 100), (3, 7), (2, 1), (1, 2), (2, 2), (4, 2), (5,), (4,), (2,)])
 @parameterize("dtype", [torch.bfloat16, torch.float16, torch.float32])
 @parameterize("fp8_format", ["e4m3", "e5m2"])

tests/test_fp8/test_fp8_fsdp_comm_hook.py

@@ -8,7 +8,7 @@ from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
 from torch.testing import assert_close
 
 from colossalai import launch
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 # example modified from https://pytorch.org/tutorials/intermediate/ddp_tutorial.html
 
@@ -28,6 +28,7 @@ class ToyModel(nn.Module):
         return self.net2(self.relu(self.net1(x)))
 
 
+@clear_cache_before_run()
 @parameterize("mode", ["grad", "params"])
 def run_model(mode):
     rank = dist.get_rank()

tests/test_fp8/test_fp8_reduce_scatter.py

@@ -6,9 +6,10 @@ from torch.testing import assert_close
 from colossalai import launch
 from colossalai.accelerator import get_accelerator
 from colossalai.quantization.fp8 import reduce_scatter_fp8
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 
 
+@clear_cache_before_run()
 @parameterize("shape", [(16, 8, 4)])
 @parameterize("scatter_dim", [0, 1, 2])
 @parameterize("dtype", [torch.bfloat16, torch.float16])