Spaces:
Running
Running
Downgrade TRL to 0.22.2 to natively bypass experimental vllm dependencies
Browse files- ultimate_sota_training.py +1 -15
ultimate_sota_training.py
CHANGED
|
@@ -66,7 +66,7 @@ def bootstrap_deps() -> None:
|
|
| 66 |
"--break-system-packages",
|
| 67 |
"httpx>=0.27.0",
|
| 68 |
"datasets>=3.4.1,<4.4.0",
|
| 69 |
-
"trl>=0.18.2,<0.
|
| 70 |
"mergekit",
|
| 71 |
"llm-blender",
|
| 72 |
"weave",
|
|
@@ -104,20 +104,6 @@ import transformers.utils.hub
|
|
| 104 |
if not hasattr(transformers.utils.hub, "TRANSFORMERS_CACHE"):
|
| 105 |
transformers.utils.hub.TRANSFORMERS_CACHE = "/tmp"
|
| 106 |
|
| 107 |
-
# CRITICAL FIX for vllm crash:
|
| 108 |
-
# Create a valid Python module on disk to satisfy both importlib and TRL's hard imports.
|
| 109 |
-
import os
|
| 110 |
-
import sys
|
| 111 |
-
vllm_dir = "/tmp/fake_vllm"
|
| 112 |
-
os.makedirs(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "pynccl"), exist_ok=True)
|
| 113 |
-
open(os.path.join(vllm_dir, "vllm", "__init__.py"), "w").close()
|
| 114 |
-
open(os.path.join(vllm_dir, "vllm", "distributed", "__init__.py"), "w").close()
|
| 115 |
-
open(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "__init__.py"), "w").close()
|
| 116 |
-
with open(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "pynccl", "__init__.py"), "w") as f:
|
| 117 |
-
f.write("class PyNcclCommunicator: pass\n")
|
| 118 |
-
if vllm_dir not in sys.path:
|
| 119 |
-
sys.path.insert(0, vllm_dir)
|
| 120 |
-
|
| 121 |
from trl import GRPOConfig, GRPOTrainer
|
| 122 |
from unsloth import FastLanguageModel
|
| 123 |
|
|
|
|
| 66 |
"--break-system-packages",
|
| 67 |
"httpx>=0.27.0",
|
| 68 |
"datasets>=3.4.1,<4.4.0",
|
| 69 |
+
"trl>=0.18.2,<=0.22.2",
|
| 70 |
"mergekit",
|
| 71 |
"llm-blender",
|
| 72 |
"weave",
|
|
|
|
| 104 |
if not hasattr(transformers.utils.hub, "TRANSFORMERS_CACHE"):
|
| 105 |
transformers.utils.hub.TRANSFORMERS_CACHE = "/tmp"
|
| 106 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
from trl import GRPOConfig, GRPOTrainer
|
| 108 |
from unsloth import FastLanguageModel
|
| 109 |
|