ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

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

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2026-03-27 14:45:12 +00:00
commit f5b3a673ee
104 changed files with 1528 additions and 1315 deletions
Download patch file Download diff file Expand all files Collapse all files

1
studio/backend/colab.py
View file

@ -82,7 +82,6 @@ def start(port: int = 8888):
from colab import start
start()
"""
import sys
logger.info("🦥 Starting Unsloth Studio...")

2
studio/backend/core/data_recipe/jobs/worker.py
View file

@ -4,8 +4,6 @@
from __future__ import annotations
import json
import structlog
import loggers
import logging
import re
import shutil

1
studio/backend/core/data_recipe/local_callable_validators.py
View file

@ -5,7 +5,6 @@ from __future__ import annotations
import json
import os
import structlog
import subprocess
from copy import deepcopy
from dataclasses import dataclass

3
studio/backend/core/export/export.py
View file

@ -8,7 +8,6 @@ Export backend - handles model exporting in various formats
import glob
import json
import structlog
from loggers import get_logger
import os
import shutil
@ -23,7 +22,7 @@ from utils.hardware import clear_gpu_cache
from utils.models import is_vision_model, get_base_model_from_lora
from utils.models.model_config import detect_audio_type
from utils.paths import ensure_dir, outputs_root, resolve_export_dir, resolve_output_dir
from utils.paths import ensure_dir, outputs_root, resolve_export_dir
from core.inference import get_inference_backend
logger = get_logger(__name__)

2
studio/backend/core/export/orchestrator.py
View file

@ -15,13 +15,11 @@ Pattern follows core/inference/orchestrator.py.
"""
import atexit
import structlog
from loggers import get_logger
import multiprocessing as mp
import queue
import threading
import time
from pathlib import Path
from typing import Any, List, Optional, Tuple
from utils.paths import outputs_root

1
studio/backend/core/export/worker.py
View file

@ -17,7 +17,6 @@ Pattern follows core/inference/worker.py and core/training/worker.py.
from __future__ import annotations
import structlog
from loggers import get_logger
import os
import sys

1
studio/backend/core/inference/audio_codecs.py
View file

@ -9,7 +9,6 @@ Supports: SNAC (Orpheus), CSM (Sesame), BiCodec (Spark), DAC (OuteTTS)
import io
import re
import wave
import structlog
from loggers import get_logger
from typing import Optional, Tuple

13
studio/backend/core/inference/inference.py
View file

@ -7,21 +7,16 @@ Core inference backend - streamlined
from unsloth import FastLanguageModel, FastVisionModel
from unsloth.chat_templates import get_chat_template
from transformers import TextStreamer
from peft import PeftModel, PeftModelForCausalLM
import json
import sys
import torch
from pathlib import Path
from typing import Optional, Union, Generator, Tuple
from utils.models import ModelConfig, get_base_model_from_lora
from utils.paths import is_model_cached
from utils.models import ModelConfig
from utils.utils import format_error_message
from utils.hardware import get_device, clear_gpu_cache, log_gpu_memory
from core.inference.audio_codecs import AudioCodecManager
from io import StringIO
import structlog
from loggers import get_logger
@ -902,7 +897,6 @@ class InferenceBackend:
try:
from utils.datasets import (
MODEL_TO_TEMPLATE_MAPPER,
get_tokenizer_chat_template,
)
model_name_lower = self.active_model_name.lower()
@ -1143,7 +1137,6 @@ class InferenceBackend:
Uses processor.apply_chat_template with audio embedded in messages (Gemma 3n pattern).
"""
import threading
import numpy as np
model_info = self.models[self.active_model_name]
model = model_info["model"]
@ -1736,7 +1729,7 @@ class InferenceBackend:
formatted_prompt = tokenizer.apply_chat_template(
chat_messages, tokenize = False, add_generation_prompt = True
)
logger.info(f"Successfully applied tokenizer's native chat template")
logger.info("Successfully applied tokenizer's native chat template")
return formatted_prompt
except Exception as e:
error_msg = str(e).lower()
@ -1745,7 +1738,7 @@ class InferenceBackend:
or "no template argument" in error_msg
):
logger.info(
f"Base model detected - no built-in chat template available, using fallback formatting"
"Base model detected - no built-in chat template available, using fallback formatting"
)
else:
logger.warning(f"Failed to apply tokenizer chat template: {e}")

7
studio/backend/core/inference/llama_cpp.py
View file

@ -12,7 +12,6 @@ import atexit
import contextlib
import json
import struct
import structlog
from loggers import get_logger
import shutil
import socket
@ -2034,7 +2033,7 @@ class LlamaCppBackend:
except httpx.ConnectError:
raise RuntimeError("Lost connection to llama-server")
except Exception as e:
except Exception:
if cancel_event is not None and cancel_event.is_set():
return
raise
@ -2629,7 +2628,7 @@ class LlamaCppBackend:
except httpx.ConnectError:
raise RuntimeError("Lost connection to llama-server")
except Exception as e:
except Exception:
if cancel_event is not None and cancel_event.is_set():
return
raise
@ -2792,7 +2791,7 @@ class LlamaCppBackend:
except httpx.ConnectError:
raise RuntimeError("Lost connection to llama-server")
except Exception as e:
except Exception:
if cancel_event is not None and cancel_event.is_set():
return
raise

2
studio/backend/core/inference/orchestrator.py
View file

@ -17,7 +17,6 @@ Pattern follows core/training/training.py.
import atexit
import base64
import structlog
from loggers import get_logger
import multiprocessing as mp
import queue
@ -25,7 +24,6 @@ import threading
import time
import uuid
from io import BytesIO
from pathlib import Path
from typing import Any, Generator, Optional, Tuple, Union
logger = get_logger(__name__)

1
studio/backend/core/inference/worker.py
View file

@ -17,7 +17,6 @@ Pattern follows core/training/worker.py.
from __future__ import annotations
import base64
import structlog
from loggers import get_logger
import os
import queue as _queue

23
studio/backend/core/training/trainer.py
View file

@ -33,22 +33,19 @@ if sys.platform in ("win32", "darwin"):
sys.path.insert(0, _compile_cache)
import torch
from utils.hardware import clear_gpu_cache, safe_num_proc, dataset_map_num_proc
from utils.hardware import clear_gpu_cache, dataset_map_num_proc
torch._dynamo.config.recompile_limit = 64
from unsloth import FastLanguageModel, FastVisionModel, is_bfloat16_supported
from unsloth.chat_templates import get_chat_template
import json
import threading
import math
import structlog
from loggers import get_logger
import time
from pathlib import Path
from typing import Optional, Callable
from dataclasses import dataclass
import pandas as pd
from datasets import Dataset, load_dataset
from utils.models import is_vision_model, detect_audio_type
@ -814,7 +811,7 @@ class UnslothTrainer:
# second attempt because the failed first call's partial
# imports clean up the stale state as a side effect.
self._source_code_retried = True
logger.info(f"\n'could not get source code' — retrying once...\n")
logger.info("\n'could not get source code' — retrying once...\n")
return self.load_model(
model_name = model_name,
max_seq_length = max_seq_length,
@ -1016,7 +1013,7 @@ class UnslothTrainer:
# Phase 2: Whisper uses FastModel.get_peft_model with task_type=None
from unsloth import FastModel
logger.info(f"Audio model (whisper) LoRA configuration:")
logger.info("Audio model (whisper) LoRA configuration:")
logger.info(f" - Target modules: {target_modules}\n")
self.model = FastModel.get_peft_model(
@ -1057,7 +1054,7 @@ class UnslothTrainer:
elif self.is_vlm:
# Vision model LoRA
logger.info(f"Vision model LoRA configuration:")
logger.info("Vision model LoRA configuration:")
logger.info(f" - Finetune vision layers: {finetune_vision_layers}")
logger.info(f" - Finetune language layers: {finetune_language_layers}")
logger.info(
@ -1085,7 +1082,7 @@ class UnslothTrainer:
)
else:
# Text model LoRA
logger.info(f"Text model LoRA configuration:")
logger.info("Text model LoRA configuration:")
logger.info(f" - Target modules: {target_modules}\n")
self.model = FastLanguageModel.get_peft_model(
@ -1114,7 +1111,6 @@ class UnslothTrainer:
except Exception as e:
import traceback
import sys
error_details = (
f"{type(e).__name__}: {str(e)}"
@ -1140,7 +1136,6 @@ class UnslothTrainer:
and strip non-TransformersKwargs params that Unsloth/PEFT inject.
"""
import types
import torch
import torch.nn as nn
from transformers.models.csm.modeling_csm import (
CsmForConditionalGeneration,
@ -1707,7 +1702,6 @@ class UnslothTrainer:
"""
import sys
import torch
import numpy as np
import torchaudio.transforms as T
import subprocess
@ -2550,7 +2544,7 @@ class UnslothTrainer:
custom_format_mapping = custom_format_mapping,
)
eval_dataset = eval_info["dataset"]
logger.info(f"Eval dataset formatted successfully\n")
logger.info("Eval dataset formatted successfully\n")
elif eval_enabled and not has_separate_eval_source:
# No separate eval source — split the already-formatted dataset
formatted_dataset = dataset_info["dataset"]
@ -3049,7 +3043,7 @@ class UnslothTrainer:
else:
# Default to warmup_steps if neither provided
config_args["warmup_steps"] = 5
logger.info(f"Using default warmup_steps: 5\n")
logger.info("Using default warmup_steps: 5\n")
# Add save_steps if specified
save_steps_val = training_args.get("save_steps", 0)
@ -3191,7 +3185,7 @@ class UnslothTrainer:
self.tokenizer, "tokenizer"
):
logger.info(
f" ⚠️ Unwrapping Processor → raw tokenizer for text-only SFTTrainer"
" ⚠️ Unwrapping Processor → raw tokenizer for text-only SFTTrainer"
)
sft_tokenizer = self.tokenizer.tokenizer
@ -3499,7 +3493,6 @@ def _ensure_deepseek_ocr_installed():
sys.path.insert(0, parent_dir)
# Try importing again
from deepseek_ocr.modeling_deepseekocr import format_messages
logger.info("DeepSeek OCR module installed successfully")
logger.info("DeepSeek OCR module installed successfully!\n")

5
studio/backend/core/training/training.py
View file

@ -19,12 +19,9 @@ import math
import multiprocessing as mp
import queue
import threading
import time
import structlog
from datetime import datetime, timezone
from loggers import get_logger
from dataclasses import dataclass, field
from pathlib import Path
from dataclasses import dataclass
from typing import Optional, Tuple, Any
import matplotlib.pyplot as plt

2
studio/backend/core/training/worker.py
View file

@ -13,7 +13,6 @@ Pattern follows core/data_recipe/jobs/worker.py.
from __future__ import annotations
import structlog
from loggers import get_logger
import os
import platform
@ -460,7 +459,6 @@ def run_training_process(
ensure_dir,
resolve_output_dir,
resolve_tensorboard_dir,
datasets_root,
)
import transformers

4
studio/backend/main.py
View file

@ -37,7 +37,7 @@ if os.getenv("ENVIRONMENT_TYPE", "production") == "production":
from fastapi import Depends, FastAPI, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from fastapi.responses import FileResponse, HTMLResponse, Response
from fastapi.responses import FileResponse, Response
from pathlib import Path
from datetime import datetime
@ -219,7 +219,7 @@ async def get_system_info():
import platform
import subprocess
import psutil
from utils.hardware import get_device, get_gpu_memory_info, DeviceType
from utils.hardware import get_gpu_memory_info
# GPU Info — query nvidia-smi for physical GPUs, filtered by
# CUDA_VISIBLE_DEVICES when set (the frontend uses this for GGUF

2
studio/backend/models/export.py
View file

@ -6,7 +6,7 @@ Pydantic schemas for Export API.
"""
from pydantic import BaseModel, Field
from typing import List, Optional, Literal, Dict, Any
from typing import Optional, Literal, Dict, Any
class LoadCheckpointRequest(BaseModel):

1
studio/backend/routes/auth.py
View file

@ -18,7 +18,6 @@ from auth import storage, hashing
from auth.authentication import (
create_access_token,
create_refresh_token,
get_current_subject,
get_current_subject_allow_password_change,
refresh_access_token,
)

1
studio/backend/routes/datasets.py
View file

@ -12,7 +12,6 @@ import sys
from pathlib import Path
from uuid import uuid4
from fastapi import APIRouter, Depends, HTTPException, UploadFile
import structlog
from loggers import get_logger
# Add backend directory to path

3
studio/backend/routes/export.py
View file

@ -7,8 +7,7 @@ Export API routes: checkpoint discovery and model export operations.
import sys
from pathlib import Path
from fastapi import APIRouter, Depends, HTTPException, Query
import structlog
from fastapi import APIRouter, Depends, HTTPException
from loggers import get_logger
# Add backend directory to path

12
studio/backend/routes/inference.py
View file

@ -13,7 +13,6 @@ from fastapi import APIRouter, Depends, HTTPException, Request
from fastapi.responses import StreamingResponse, JSONResponse
from typing import Optional
import json
import structlog
from loggers import get_logger
import asyncio
import threading
@ -82,8 +81,6 @@ from models.inference import (
)
from auth.authentication import get_current_subject
import io
import wave
import base64
import numpy as np
@ -328,14 +325,14 @@ async def load_model(
training_method = adapter_cfg.get("unsloth_training_method")
if training_method == "lora" and load_in_4bit:
logger.info(
f"adapter_config.json says unsloth_training_method='lora' — "
f"setting load_in_4bit=False to match 16-bit training"
"adapter_config.json says unsloth_training_method='lora' — "
"setting load_in_4bit=False to match 16-bit training"
)
load_in_4bit = False
elif training_method == "qlora" and not load_in_4bit:
logger.info(
f"adapter_config.json says unsloth_training_method='qlora' — "
f"setting load_in_4bit=True to match QLoRA training"
"adapter_config.json says unsloth_training_method='qlora' — "
"setting load_in_4bit=True to match QLoRA training"
)
load_in_4bit = True
elif training_method:
@ -752,7 +749,6 @@ async def generate_audio(
def _decode_audio_base64(b64: str) -> np.ndarray:
"""Decode base64 audio (any format) → float32 numpy array at 16kHz."""
import torch
import torchaudio
import tempfile
import os

2
studio/backend/routes/models.py
View file

@ -10,7 +10,6 @@ import sys
from pathlib import Path
from fastapi import APIRouter, Body, Depends, HTTPException, Query
from typing import List, Optional
import structlog
from loggers import get_logger
import re as _re
@ -775,7 +774,6 @@ async def get_gguf_variants(
# case-insensitive match.
cached_bytes_by_quant: dict[str, int] = {}
try:
import re as _re
from huggingface_hub import constants as hf_constants
# Sanitize repo_id: must be "owner/name" with safe chars only

3
studio/backend/routes/training.py
View file

@ -9,8 +9,7 @@ import sys
from pathlib import Path
from fastapi import APIRouter, Depends, HTTPException, Request
from fastapi.responses import StreamingResponse
from typing import Dict, Optional, Any
import structlog
from typing import Optional, Any
from loggers import get_logger
import asyncio
from datetime import datetime

1
studio/backend/tests/test_transformers_version.py
View file

@ -4,7 +4,6 @@
"""Tests for transformers version detection with local checkpoint fallbacks."""
import json
import pytest
from pathlib import Path
from unittest.mock import patch

5
studio/backend/tests/test_utils.py
View file

@ -17,7 +17,6 @@ Run with:
python -m pytest tests/test_utils.py -v
"""
import platform
from unittest.mock import patch, MagicMock
import pytest
@ -295,8 +294,6 @@ class TestLogGpuMemory:
"utilization_pct": 12.5,
"free_gb": 14.0,
}
import structlog
from loggers import get_logger
with (
patch(
@ -312,8 +309,6 @@ class TestLogGpuMemory:
def test_logs_cpu_fallback_when_no_gpu(self, caplog):
fake_info = {"available": False, "backend": "cpu"}
import structlog
from loggers import get_logger
with (
patch(

1
studio/backend/utils/cache_cleanup.py
View file

@ -12,7 +12,6 @@ by spawned subprocesses.
"""
import shutil
import structlog
from loggers import get_logger
from pathlib import Path
from typing import List, Optional

111
studio/backend/utils/datasets/chat_templates.py
View file

@ -8,14 +8,13 @@ This module contains functions for applying chat templates to datasets
and generating dataset info summaries.
"""
from .format_detection import detect_dataset_format, detect_multimodal_dataset, detect_custom_format_heuristic
from .format_detection import detect_custom_format_heuristic
from .model_mappings import MODEL_TO_TEMPLATE_MAPPER
from loggers import get_logger
logger = get_logger(__name__)
DEFAULT_ALPACA_TEMPLATE = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
@ -63,18 +62,21 @@ def get_tokenizer_chat_template(tokenizer, model_name):
)
except Exception as e:
logger.info(f"⚠️ Failed to apply Unsloth template '{matched_template}': {e}")
logger.info(f" Falling back to tokenizer's default chat template")
logger.info(" Falling back to tokenizer's default chat template")
else:
# Check if tokenizer actually has a chat_template set
has_chat_template = (
hasattr(tokenizer, 'chat_template')
and tokenizer.chat_template is not None
hasattr(tokenizer, "chat_template") and tokenizer.chat_template is not None
)
if has_chat_template:
logger.info(f"📝 Using tokenizer's own chat template (no Unsloth template match)")
logger.info(
"📝 Using tokenizer's own chat template (no Unsloth template match)"
)
else:
# Base model with no chat template — apply default ChatML
logger.info(f"📝 No chat template found — applying default ChatML template (base model)")
logger.info(
"📝 No chat template found — applying default ChatML template (base model)"
)
try:
tokenizer = get_chat_template(
tokenizer,
@ -82,7 +84,7 @@ def get_tokenizer_chat_template(tokenizer, model_name):
)
except Exception as e:
logger.info(f"⚠️ Failed to apply default ChatML template: {e}")
logger.info(f" Falling back to tokenizer as-is")
logger.info(" Falling back to tokenizer as-is")
return tokenizer
@ -99,7 +101,7 @@ def get_dataset_info_summary(dataset_info):
"sharegpt": "ShareGPT format (needs standardization)",
"chatml_messages": "ChatML format (messages column) - OpenAI compatible",
"chatml_conversations": "ChatML format (conversations column) - HuggingFace standard",
"unknown": "Unknown format"
"unknown": "Unknown format",
}
return {
@ -110,7 +112,8 @@ def get_dataset_info_summary(dataset_info):
"chat_column": dataset_info["chat_column"],
"is_standardized": dataset_info["is_standardized"],
"warnings": dataset_info.get("warnings", []),
"ready_for_training": dataset_info["is_standardized"] and final_format != "unknown"
"ready_for_training": dataset_info["is_standardized"]
and final_format != "unknown",
}
@ -154,7 +157,7 @@ def apply_chat_template_to_dataset(
# Get EOS token if needed
eos_token = ""
if add_eos_token:
if hasattr(tokenizer, 'eos_token') and tokenizer.eos_token:
if hasattr(tokenizer, "eos_token") and tokenizer.eos_token:
eos_token = tokenizer.eos_token
else:
warnings.append("add_eos_token=True but tokenizer has no eos_token")
@ -167,14 +170,16 @@ def apply_chat_template_to_dataset(
if not dataset_info.get("auto_detection_attempted", False):
custom_format_mapping = detect_custom_format_heuristic(dataset)
if custom_format_mapping:
warnings.append(f"Auto-detected column mapping: {custom_format_mapping}")
warnings.append(
f"Auto-detected column mapping: {custom_format_mapping}"
)
else:
errors.append("Could not auto-detect format mapping")
return {
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}
else:
# Already failed once in format_dataset, don't retry
@ -186,7 +191,7 @@ def apply_chat_template_to_dataset(
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}
if custom_format_mapping:
@ -209,7 +214,7 @@ def apply_chat_template_to_dataset(
for i in range(num_examples):
convo = []
role_order = ['system', 'user', 'assistant']
role_order = ["system", "user", "assistant"]
for target_role in role_order:
for col_name, role in custom_format_mapping.items():
@ -218,11 +223,18 @@ def apply_chat_template_to_dataset(
if is_user_provided:
# User explicitly mapped - include even if empty
convo.append({"role": role, "content": str(content) if content else ""})
convo.append(
{
"role": role,
"content": str(content) if content else "",
}
)
else:
# Auto-detected - skip empty
if content and str(content).strip():
convo.append({"role": role, "content": str(content)})
convo.append(
{"role": role, "content": str(content)}
)
conversations.append(convo)
@ -232,31 +244,35 @@ def apply_chat_template_to_dataset(
return result
try:
dataset = dataset.map(_apply_custom_mapping, batched = True, batch_size = batch_size)
dataset = dataset.map(
_apply_custom_mapping, batched = True, batch_size = batch_size
)
# Update to use conversations format
final_format = "chatml_conversations"
chat_column = "conversations"
is_standardized = True
warnings.append("Successfully converted to ChatML format via custom mapping")
warnings.append(
"Successfully converted to ChatML format via custom mapping"
)
except Exception as e:
errors.append(f"Custom format mapping failed: {e}")
return {
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}
# ALPACA FORMAT
if final_format == "alpaca":
# Set alpaca chat template on tokenizer for saving (if not already set)
# This ensures the template is saved with the model for inference
if not (hasattr(tokenizer, 'chat_template') and tokenizer.chat_template):
if not (hasattr(tokenizer, "chat_template") and tokenizer.chat_template):
try:
from unsloth.chat_templates import get_chat_template
tokenizer = get_chat_template(tokenizer, chat_template = "alpaca")
logger.info(f"📝 Set alpaca chat template on tokenizer for model saving")
logger.info("📝 Set alpaca chat template on tokenizer for model saving")
except Exception as e:
logger.info(f"⚠️ Could not set alpaca template on tokenizer: {e}")
@ -266,12 +282,16 @@ def apply_chat_template_to_dataset(
for i in range(len(examples["instruction"])):
fields = {
"instruction": examples["instruction"][i],
"input": examples.get("input", [""] * len(examples["instruction"]))[i],
"output": examples["output"][i]
"input": examples.get("input", [""] * len(examples["instruction"]))[
i
],
"output": examples["output"][i],
}
try:
text = DEFAULT_ALPACA_TEMPLATE.format(fields["instruction"], fields["input"], fields["output"])
text = DEFAULT_ALPACA_TEMPLATE.format(
fields["instruction"], fields["input"], fields["output"]
)
text += eos_token
texts.append(text)
except KeyError as e:
@ -284,24 +304,26 @@ def apply_chat_template_to_dataset(
try:
dataset_map_kwargs = {
'batched': True,
'batch_size': batch_size,
"batched": True,
"batch_size": batch_size,
}
try:
from torch.utils.data import IterableDataset
_is_torch_iterable = isinstance(dataset, IterableDataset)
except ImportError:
_is_torch_iterable = False
if not _is_torch_iterable:
from utils.hardware import dataset_map_num_proc
if num_proc is None or type(num_proc) is not int:
num_proc = dataset_map_num_proc()
else:
num_proc = dataset_map_num_proc(num_proc)
dataset_map_kwargs['num_proc'] = num_proc
dataset_map_kwargs['desc'] = "Applying template to Alpaca format"
dataset_map_kwargs["num_proc"] = num_proc
dataset_map_kwargs["desc"] = "Applying template to Alpaca format"
formatted_dataset = dataset.map(formatted_fn, **dataset_map_kwargs)
@ -309,7 +331,7 @@ def apply_chat_template_to_dataset(
"dataset": formatted_dataset,
"success": True,
"warnings": warnings,
"errors": errors
"errors": errors,
}
except Exception as e:
errors.append(f"Failed to format Alpaca dataset: {e}")
@ -317,12 +339,11 @@ def apply_chat_template_to_dataset(
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}
# CHATML FORMATS
elif final_format in ["chatml_messages", "chatml_conversations"]:
if not is_standardized:
warnings.append("Dataset may not be fully standardized")
@ -337,13 +358,11 @@ def apply_chat_template_to_dataset(
for convo in convos:
try:
text = tokenizer.apply_chat_template(
convo,
tokenize = False,
add_generation_prompt = False
convo, tokenize = False, add_generation_prompt = False
)
if remove_bos_prefix:
text = text.removeprefix('<bos>')
text = text.removeprefix("<bos>")
text += eos_token
texts.append(text)
@ -357,23 +376,25 @@ def apply_chat_template_to_dataset(
try:
try:
from torch.utils.data import IterableDataset
_is_torch_iterable = isinstance(dataset, IterableDataset)
except ImportError:
_is_torch_iterable = False
dataset_map_kwargs = {
'batched': True,
'batch_size': batch_size,
"batched": True,
"batch_size": batch_size,
}
if not _is_torch_iterable:
from utils.hardware import dataset_map_num_proc
if num_proc is None or type(num_proc) is not int:
num_proc = dataset_map_num_proc()
else:
num_proc = dataset_map_num_proc(num_proc)
dataset_map_kwargs['num_proc'] = num_proc
dataset_map_kwargs['desc'] = f"Applying chat template to {final_format}"
dataset_map_kwargs["num_proc"] = num_proc
dataset_map_kwargs["desc"] = f"Applying chat template to {final_format}"
# Monitor tqdm progress from dataset.map() and relay to callback
_tqdm_monitor_stop = None
@ -411,7 +432,7 @@ def apply_chat_template_to_dataset(
"dataset": formatted_dataset,
"success": True,
"warnings": warnings,
"errors": errors
"errors": errors,
}
except Exception as e:
errors.append(f"Failed to format ChatML dataset: {e}")
@ -419,7 +440,7 @@ def apply_chat_template_to_dataset(
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}
# UNKNOWN FORMAT
@ -432,5 +453,5 @@ def apply_chat_template_to_dataset(
"dataset": dataset,
"success": False,
"warnings": warnings,
"errors": errors
"errors": errors,
}

3
studio/backend/utils/datasets/data_collators.py
View file

@ -9,7 +9,7 @@ particularly for VLM/OCR processing.
"""
from dataclasses import dataclass
from typing import Any, List, Optional, Union
from typing import Any, List
from loggers import get_logger
logger = get_logger(__name__)
@ -75,7 +75,6 @@ class DeepSeekOCRDataCollator:
Returns:
dict with input_ids, attention_mask, labels, pixel_values, etc.
"""
from PIL import Image
# Extract messages and images
all_messages = []

11
studio/backend/utils/datasets/dataset_utils.py
View file

@ -38,12 +38,7 @@ from .format_conversion import (
from .chat_templates import (
apply_chat_template_to_dataset,
get_dataset_info_summary,
get_tokenizer_chat_template,
DEFAULT_ALPACA_TEMPLATE,
)
from .vlm_processing import generate_smart_vlm_instruction
from .data_collators import DeepSeekOCRDataCollator, VLMDataCollator
from .model_mappings import TEMPLATE_TO_MODEL_MAPPER
from loggers import get_logger
logger = get_logger(__name__)
@ -701,7 +696,7 @@ def format_dataset(
}
else:
warnings.append(f"Cannot convert unknown format to Alpaca")
warnings.append("Cannot convert unknown format to Alpaca")
return {
"dataset": dataset,
"detected_format": "unknown",
@ -766,7 +761,7 @@ def format_dataset(
}
else:
warnings.append(f"Unknown format, attempting standardization")
warnings.append("Unknown format, attempting standardization")
if detected["chat_column"]:
try:
standardized = standardize_chat_format(
@ -914,7 +909,7 @@ def format_and_template_dataset(
f"falling back to auto-detection"
)
logger.info(
f"⚠️ User VLM mapping failed, falling back to auto-detection..."
"⚠️ User VLM mapping failed, falling back to auto-detection..."
)
custom_format_mapping = None # clear so auto-detection runs below
else:

1
studio/backend/utils/datasets/format_conversion.py
View file

@ -41,7 +41,6 @@ def standardize_chat_format(
"""
import collections
import itertools
from datasets import IterableDataset
# Check if vision tokenizer is used
is_vlm = False

2
studio/backend/utils/datasets/format_detection.py
View file

@ -229,7 +229,7 @@ def detect_custom_format_heuristic(dataset):
if prefix in ["generation", "pass", "inference"]:
return True
if len(col_lower) <= 2 and not col_lower in ["qa", "q", "a"]:
if len(col_lower) <= 2 and col_lower not in ["qa", "q", "a"]:
return True
return False

3
studio/backend/utils/datasets/llm_assist.py
View file

@ -19,7 +19,6 @@ import os
import re
import textwrap
import time
from itertools import islice
from typing import Any, Optional
from loggers import get_logger
@ -105,7 +104,7 @@ def precache_helper_gguf():
finally:
try:
enable_progress_bars()
except Exception as e:
except Exception:
pass

2
studio/backend/utils/hardware/hardware.py
View file

@ -17,7 +17,6 @@ Usage:
"""
import platform
import structlog
from loggers import get_logger
from enum import Enum
from typing import Optional, Dict, Any
@ -183,7 +182,6 @@ def get_gpu_memory_info() -> Dict[str, Any]:
# ---- MLX path (Apple Silicon) ----
if device == DeviceType.MLX:
try:
import mlx.core as mx
import psutil
# MLX uses unified memory — report system memory as the pool

1
studio/backend/utils/inference/inference_config.py
View file

@ -13,7 +13,6 @@ from pathlib import Path
from typing import Dict, Any, Optional
import json
import yaml
import structlog
from loggers import get_logger
from utils.models.model_config import load_model_defaults

1
studio/backend/utils/models/checkpoints.py
View file

@ -6,7 +6,6 @@ Checkpoint scanning utilities for discovering training runs and their checkpoint
"""
import json
import structlog
from loggers import get_logger
from pathlib import Path
from typing import List, Optional, Tuple

1
studio/backend/utils/models/model_config.py
View file

@ -18,7 +18,6 @@ from utils.paths import (
resolve_export_dir,
)
from utils.utils import without_hf_auth
import structlog
from loggers import get_logger
import os
import subprocess

1
studio/backend/utils/paths/path_utils.py
View file

@ -8,7 +8,6 @@ Path utilities for model and dataset handling
import os
from pathlib import Path
from typing import Optional
import structlog
from loggers import get_logger
logger = get_logger(__name__)

1
studio/backend/utils/transformers_version.py
View file

@ -23,7 +23,6 @@ Strategy:
import importlib
import json
import structlog
from loggers import get_logger
import os
import shutil

1
studio/backend/utils/utils.py
View file

@ -6,7 +6,6 @@ Shared backend utilities
"""
import os
import structlog
from loggers import get_logger
from contextlib import contextmanager
from pathlib import Path

2
studio/install_python_stack.py
View file

@ -353,7 +353,7 @@ def pip_install(
)
if result.returncode == 0:
return
print(_red(f" uv failed, falling back to pip..."))
print(_red(" uv failed, falling back to pip..."))
if result.stdout:
print(result.stdout.decode(errors = "replace"))

1
tests/python/test_cross_platform_parity.py
View file

@ -5,7 +5,6 @@ from __future__ import annotations
import re
from pathlib import Path
import pytest
REPO_ROOT = Path(__file__).resolve().parents[2]
INSTALL_SH = REPO_ROOT / "install.sh"

1
tests/python/test_e2e_no_torch_sandbox.py
View file

@ -21,7 +21,6 @@ from __future__ import annotations
import os
import shutil
import signal
import subprocess
import sys
import textwrap

2
tests/python/test_install_python_stack.py
View file

@ -2,13 +2,11 @@
from __future__ import annotations
import importlib
import os
import sys
from pathlib import Path
from unittest import mock
import pytest
# Add the studio directory so we can import install_python_stack
STUDIO_DIR = Path(__file__).resolve().parents[2] / "studio"

1
tests/python/test_no_torch_filtering.py
View file

@ -10,7 +10,6 @@ Covers:
from __future__ import annotations
import importlib
import os
import re
import subprocess

1
tests/python/test_studio_import_no_torch.py
View file

@ -15,7 +15,6 @@ import ast
import os
import shutil
import subprocess
import sys
import tempfile
import textwrap
from pathlib import Path

2
tests/saving/language_models/test_merge_4bit_validation.py
View file

@ -1,6 +1,6 @@
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template
from trl import SFTTrainer, SFTConfig
from trl import SFTTrainer
from transformers import DataCollatorForSeq2Seq, TrainingArguments
from datasets import load_dataset
import torch

1
tests/saving/language_models/test_save_merged_grpo_model.py
View file

@ -11,7 +11,6 @@ import sys
from pathlib import Path
import multiprocessing as mp
import gc
from multiprocessing import Queue
REPO_ROOT = Path(__file__).parents[3]
sys.path.insert(0, str(REPO_ROOT))

2
tests/saving/non_peft/test_mistral_non_peft.py
View file

@ -1,6 +1,4 @@
from unsloth import FastLanguageModel
from transformers import AutoModelForCausalLM
from peft import PeftModel
from pathlib import Path
import sys
import warnings

5
tests/saving/non_peft/test_whisper_non_peft.py
View file

@ -1,6 +1,5 @@
from unsloth import FastLanguageModel, FastModel
from transformers import AutoModelForCausalLM, WhisperForConditionalGeneration
from peft import PeftModel
from unsloth import FastModel
from transformers import WhisperForConditionalGeneration
from pathlib import Path
import sys
import warnings

18
tests/saving/vision_models/test_index_file_sharded_model.py
View file

@ -28,7 +28,7 @@ train_dataset = dataset.select(range(2000))
# To select the next 200 examples for evaluation
eval_dataset = dataset.select(range(2000, 2200))
print(f"✅ Dataset loaded successfully!")
print("✅ Dataset loaded successfully!")
print(f" 📈 Training samples: {len(train_dataset)}")
print(f" 📊 Evaluation samples: {len(eval_dataset)}")
@ -110,10 +110,10 @@ try:
loftq_config = None, # And LoftQ
)
print("✅ LoRA configuration applied successfully!")
print(f" 🎯 LoRA rank (r): 16")
print(f" 📊 LoRA alpha: 32")
print(f" 🔍 Vision layers: Enabled")
print(f" 💬 Language layers: Enabled")
print(" 🎯 LoRA rank (r): 16")
print(" 📊 LoRA alpha: 32")
print(" 🔍 Vision layers: Enabled")
print(" 💬 Language layers: Enabled")
except Exception as e:
print(f"❌ Failed to apply LoRA configuration: {e}")
raise
@ -165,10 +165,10 @@ try:
),
)
print("✅ Trainer setup completed!")
print(f" 📦 Batch size: 2")
print(f" 🔄 Gradient accumulation steps: 4")
print(f" 📈 Max training steps: 10")
print(f" 🎯 Learning rate: 2e-4")
print(" 📦 Batch size: 2")
print(" 🔄 Gradient accumulation steps: 4")
print(" 📈 Max training steps: 10")
print(" 🎯 Learning rate: 2e-4")
print(f" 💾 Precision: {'BF16' if is_bf16_supported() else 'FP16'}")
except Exception as e:
print(f"❌ Failed to setup trainer: {e}")

18
tests/saving/vision_models/test_push_to_hub_merged.py
View file

@ -29,7 +29,7 @@ train_dataset = dataset.select(range(2000))
# To select the next 200 examples for evaluation
eval_dataset = dataset.select(range(2000, 2200))
print(f"✅ Dataset loaded successfully!")
print("✅ Dataset loaded successfully!")
print(f" 📈 Training samples: {len(train_dataset)}")
print(f" 📊 Evaluation samples: {len(eval_dataset)}")
@ -111,10 +111,10 @@ try:
loftq_config = None, # And LoftQ
)
print("✅ LoRA configuration applied successfully!")
print(f" 🎯 LoRA rank (r): 16")
print(f" 📊 LoRA alpha: 32")
print(f" 🔍 Vision layers: Enabled")
print(f" 💬 Language layers: Enabled")
print(" 🎯 LoRA rank (r): 16")
print(" 📊 LoRA alpha: 32")
print(" 🔍 Vision layers: Enabled")
print(" 💬 Language layers: Enabled")
except Exception as e:
print(f"❌ Failed to apply LoRA configuration: {e}")
raise
@ -166,10 +166,10 @@ try:
),
)
print("✅ Trainer setup completed!")
print(f" 📦 Batch size: 2")
print(f" 🔄 Gradient accumulation steps: 4")
print(f" 📈 Max training steps: 10")
print(f" 🎯 Learning rate: 2e-4")
print(" 📦 Batch size: 2")
print(" 🔄 Gradient accumulation steps: 4")
print(" 📈 Max training steps: 10")
print(" 🎯 Learning rate: 2e-4")
print(f" 💾 Precision: {'BF16' if is_bf16_supported() else 'FP16'}")
except Exception as e:
print(f"❌ Failed to setup trainer: {e}")

10
tests/saving/vision_models/test_save_merge_qwen2.5vl32B_model_ocr_benchmark.py
View file

@ -2,9 +2,6 @@
from unsloth import FastVisionModel
import torch
from qwen_vl_utils import process_vision_info
import os
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig
@ -20,7 +17,6 @@ from tests.utils.ocr_eval import OCRModelEvaluator
## Dataset Preparation
from datasets import load_dataset
dataset = load_dataset("lbourdois/OCR-liboaccn-OPUS-MIT-5M-clean", "en", split = "train")
# To select the first 2000 examples
@ -66,12 +62,6 @@ train_dataset = [format_data(sample) for sample in train_dataset]
eval_dataset = [format_data(sample) for sample in eval_dataset]
## Setup OCR main evaluation function and helpers
import os
import torch
from tqdm import tqdm
import pandas as pd
from jiwer import wer, cer
from qwen_vl_utils import process_vision_info
#
ocr_evaluator = OCRModelEvaluator()

10
tests/saving/vision_models/test_save_merge_vision_model_ocr_benchmark.py
View file

@ -2,9 +2,6 @@
from unsloth import FastVisionModel
import torch
from qwen_vl_utils import process_vision_info
import os
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig
@ -20,7 +17,6 @@ from tests.utils.ocr_eval import OCRModelEvaluator
## Dataset Preparation
from datasets import load_dataset
dataset = load_dataset("lbourdois/OCR-liboaccn-OPUS-MIT-5M-clean", "en", split = "train")
# To select the first 2000 examples
@ -66,12 +62,6 @@ train_dataset = [format_data(sample) for sample in train_dataset]
eval_dataset = [format_data(sample) for sample in eval_dataset]
## Setup OCR main evaluation function and helpers
import os
import torch
from tqdm import tqdm
import pandas as pd
from jiwer import wer, cer
from qwen_vl_utils import process_vision_info
#
ocr_evaluator = OCRModelEvaluator()

1
tests/studio/install/test_install_llama_prebuilt_logic.py
View file

@ -1,6 +1,5 @@
import importlib.util
import io
import json
import os
import sys
import tarfile

2
tests/studio/install/test_pr4562_bugfixes.py
View file

@ -14,13 +14,11 @@ Run: pytest tests/studio/install/test_pr4562_bugfixes.py -v
"""
import importlib.util
import json
import os
import subprocess
import sys
import textwrap
from pathlib import Path
from unittest.mock import patch
import pytest

1
tests/test_raw_text.py
View file

@ -7,7 +7,6 @@ Tests basic functionality without heavy dependencies.
import sys
import os
import tempfile
from pathlib import Path
import importlib.util

12
tests/utils/aime_eval.py
View file

@ -207,7 +207,7 @@ def evaluate_model_aime(
print(f"\n{'='*70}")
print(f"🧮 AIME EVALUATION - {model_type.upper()} MODEL")
print(f"Combined Dataset: test2024 + test2025-I + test2025-II")
print("Combined Dataset: test2024 + test2025-I + test2025-II")
print(f"{'='*70}")
# Load combined AIME dataset
@ -244,7 +244,7 @@ def evaluate_model_aime(
seed = seed,
)
print(f"\n🔧 Configuration:")
print("\n🔧 Configuration:")
print(f" Temperature: {temperature}")
print(f" Samples per question: {n_sampling}")
print(f" Max tokens: {max_tokens}")
@ -421,28 +421,28 @@ def evaluate_model_aime(
print(f"📊 AIME EVALUATION RESULTS - {model_type.upper()}")
print(f"{'='*70}")
print(f"\n🎯 Overall Performance:")
print("\n🎯 Overall Performance:")
print(f" Total problems: {total_problems:>6}")
print(
f" Correct answers: {correct_answers:>6}/{total_problems} ({accuracy:>5.1f}%)"
)
print(f" Pass@{n_sampling}: {pass_at_k:>10.1f}%")
print(f"\n📈 Performance by Dataset:")
print("\n📈 Performance by Dataset:")
for source, stats in source_stats.items():
source_acc = source_accuracies[source]
print(
f" {source:>12}: {stats['correct']:>3}/{stats['total']:>3} ({source_acc:>5.1f}%)"
)
print(f"\n🔧 Configuration:")
print("\n🔧 Configuration:")
print(f" Temperature: {temperature}")
print(f" Samples per problem: {n_sampling}")
print(f" Max tokens: {max_tokens}")
print(f" Top-p: {top_p}")
print(f" Seed: {seed}")
print(f"\n📝 Token Statistics:")
print("\n📝 Token Statistics:")
print(f" Avg input tokens: {results['avg_input_tokens']:>10.1f}")
print(f" Avg output tokens: {results['avg_output_tokens']:>10.1f}")
print(f" Max input tokens: {results['max_input_tokens']:>10}")

5
tests/utils/hf_utils.py
View file

@ -12,9 +12,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from contextlib import contextmanager, nullcontext
from typing import Callable, Optional
from contextlib import nullcontext
from typing import Callable
import bitsandbytes as bnb
import torch

2
tests/utils/ocr_eval.py
View file

@ -348,7 +348,7 @@ class OCRModelEvaluator:
plt.savefig("ocr_model_comparison.png")
plt.show()
print(f"\nVisualization saved to ocr_model_comparison.png")
print("\nVisualization saved to ocr_model_comparison.png")
def get_comparison_results(self) -> Dict[str, Dict[str, float]]:
"""Get the current comparison results."""

8
tests/utils/os_utils.py
View file

@ -98,10 +98,10 @@ def require_package(package_name, executable_name = None):
for pm_name, cmd in install_commands.items():
print(f" {pm_name}: {cmd}")
print(f"\nAlternatively, install with conda:")
print("\nAlternatively, install with conda:")
print(f" conda install -c conda-forge {package_name}")
print(f"\nPlease install the required package and run the script again.")
print("\nPlease install the required package and run the script again.")
sys.exit(1)
@ -120,9 +120,9 @@ def require_python_package(package_name, import_name = None, pip_name = None):
print(f"❌ Error: Python package '{package_name}' is not installed")
print(f"\nPlease install {package_name} using pip:")
print(f" pip install {pip_name}")
print(f" # or with conda:")
print(" # or with conda:")
print(f" conda install {pip_name}")
print(f"\nAfter installation, run this script again.")
print("\nAfter installation, run this script again.")
sys.exit(1)
else:
print(f"✓ Python package '{package_name}' is installed")

1
tests/utils/test_q_galore.py
View file

@ -14,7 +14,6 @@
#
# Tests for Q-GaLore integration (unsloth/optimizers/).
import pytest
import sys
import os
import torch

2
unsloth-cli.py
View file

@ -101,8 +101,6 @@ def run(args):
return {"text": texts}
def load_dataset_smart(args):
from transformers.utils import strtobool
if args.raw_text_file:
# Use raw text loader
loader = RawTextDataLoader(tokenizer, args.chunk_size, args.stride)

12
unsloth/__init__.py
View file

@ -12,9 +12,15 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import warnings, importlib, sys
import warnings
import importlib
import sys
from packaging.version import Version
import os, re, subprocess, inspect, functools
import os
import re
import subprocess
import inspect
import functools
import numpy as np
# Log Unsloth is being used
@ -105,7 +111,7 @@ try:
import unsloth_zoo
except PackageNotFoundError:
raise ImportError(
f"Unsloth: Please install unsloth_zoo via `pip install unsloth_zoo` then retry!"
"Unsloth: Please install unsloth_zoo via `pip install unsloth_zoo` then retry!"
)
except:
raise

74
unsloth/_auto_install.py
View file

@ -12,32 +12,60 @@
# See the License for the specific language governing permissions and
# limitations under the License.
try: import torch
except: raise ImportError('Install torch via `pip install torch`')
try:
import torch
except:
raise ImportError("Install torch via `pip install torch`")
from packaging.version import Version as V
import re
v = V(re.match(r"[0-9\.]{3,}", torch.__version__).group(0))
cuda = str(torch.version.cuda)
is_ampere = torch.cuda.get_device_capability()[0] >= 8
USE_ABI = torch._C._GLIBCXX_USE_CXX11_ABI
if cuda not in ("11.8", "12.1", "12.4", "12.6", "12.8", "13.0"): raise RuntimeError(f"CUDA = {cuda} not supported!")
if v <= V('2.1.0'): raise RuntimeError(f"Torch = {v} too old!")
elif v <= V('2.1.1'): x = 'cu{}{}-torch211'
elif v <= V('2.1.2'): x = 'cu{}{}-torch212'
elif v < V('2.3.0'): x = 'cu{}{}-torch220'
elif v < V('2.4.0'): x = 'cu{}{}-torch230'
elif v < V('2.5.0'): x = 'cu{}{}-torch240'
elif v < V('2.5.1'): x = 'cu{}{}-torch250'
elif v <= V('2.5.1'): x = 'cu{}{}-torch251'
elif v < V('2.7.0'): x = 'cu{}{}-torch260'
elif v < V('2.7.9'): x = 'cu{}{}-torch270'
elif v < V('2.8.0'): x = 'cu{}{}-torch271'
elif v < V('2.8.9'): x = 'cu{}{}-torch280'
elif v < V('2.9.1'): x = 'cu{}{}-torch290'
elif v < V('2.9.2'): x = 'cu{}{}-torch291'
elif v < V('2.10.1'): x = 'cu{}{}-torch2100'
else: raise RuntimeError(f"Torch = {v} too new!")
if v > V('2.6.9') and cuda not in ("11.8", "12.6", "12.8", "13.0"): raise RuntimeError(f"CUDA = {cuda} not supported!")
if v >= V('2.10.0') and cuda not in ("12.6", "12.8", "13.0"): raise RuntimeError(f"Torch 2.10 requires CUDA 12.6, 12.8, or 13.0! Got CUDA = {cuda}")
x = x.format(cuda.replace(".", ""), "-ampere" if False else "") # is_ampere is broken due to flash-attn
print(f'pip install --upgrade pip && pip install --no-deps git+https://github.com/unslothai/unsloth-zoo.git && pip install "unsloth[{x}] @ git+https://github.com/unslothai/unsloth.git" --no-build-isolation')
if cuda not in ("11.8", "12.1", "12.4", "12.6", "12.8", "13.0"):
raise RuntimeError(f"CUDA = {cuda} not supported!")
if v <= V("2.1.0"):
raise RuntimeError(f"Torch = {v} too old!")
elif v <= V("2.1.1"):
x = "cu{}{}-torch211"
elif v <= V("2.1.2"):
x = "cu{}{}-torch212"
elif v < V("2.3.0"):
x = "cu{}{}-torch220"
elif v < V("2.4.0"):
x = "cu{}{}-torch230"
elif v < V("2.5.0"):
x = "cu{}{}-torch240"
elif v < V("2.5.1"):
x = "cu{}{}-torch250"
elif v <= V("2.5.1"):
x = "cu{}{}-torch251"
elif v < V("2.7.0"):
x = "cu{}{}-torch260"
elif v < V("2.7.9"):
x = "cu{}{}-torch270"
elif v < V("2.8.0"):
x = "cu{}{}-torch271"
elif v < V("2.8.9"):
x = "cu{}{}-torch280"
elif v < V("2.9.1"):
x = "cu{}{}-torch290"
elif v < V("2.9.2"):
x = "cu{}{}-torch291"
elif v < V("2.10.1"):
x = "cu{}{}-torch2100"
else:
raise RuntimeError(f"Torch = {v} too new!")
if v > V("2.6.9") and cuda not in ("11.8", "12.6", "12.8", "13.0"):
raise RuntimeError(f"CUDA = {cuda} not supported!")
if v >= V("2.10.0") and cuda not in ("12.6", "12.8", "13.0"):
raise RuntimeError(
f"Torch 2.10 requires CUDA 12.6, 12.8, or 13.0! Got CUDA = {cuda}"
)
x = x.format(
cuda.replace(".", ""), "-ampere" if False else ""
) # is_ampere is broken due to flash-attn
print(
f'pip install --upgrade pip && pip install --no-deps git+https://github.com/unslothai/unsloth-zoo.git && pip install "unsloth[{x}] @ git+https://github.com/unslothai/unsloth.git" --no-build-isolation'
)

1506
unsloth/chat_templates.py
View file

File diff suppressed because it is too large Load diff

2
unsloth/dataprep/raw_text.py
View file

@ -12,11 +12,9 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import re
import json
import csv
from typing import List, Dict, Any, Union, Optional
from datasets import Dataset
from pathlib import Path

1
unsloth/dataprep/synthetic.py
View file

@ -25,7 +25,6 @@ os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
import requests
import torch
import gc
import time
import re
from unsloth_zoo.log import logger
import numpy as np

4
unsloth/import_fixes.py
View file

@ -75,7 +75,7 @@ class HideLoggingMessage(logging.Filter):
self.text = text
def filter(self, x):
return not (self.text in x.getMessage())
return self.text not in x.getMessage()
class HidePrintMessage:
@ -1310,7 +1310,7 @@ def disable_broken_wandb():
return # wandb not installed, nothing to do
try:
import wandb
pass
except Exception:
# wandb is installed but broken - patch all checkers to skip it
logger.info(

2
unsloth/kernels/cross_entropy_loss.py
View file

@ -23,12 +23,10 @@ from .utils import (
torch_gpu_device,
is_cdna,
)
from transformers.models.llama.modeling_llama import logger
from unsloth_zoo.utils import Version
from unsloth_zoo.loss_utils import (
patch_loss_functions as _patch_loss_functions,
post_patch_loss_function,
)

1
unsloth/kernels/fast_lora.py
View file

@ -18,7 +18,6 @@ from .utils import (
fast_dequantize,
QUANT_STATE,
get_lora_parameters,
get_lora_parameters_bias,
matmul_lora,
torch_amp_custom_fwd,
torch_amp_custom_bwd,

5
unsloth/kernels/flex_attention.py
View file

@ -13,8 +13,6 @@
# limitations under the License.
import torch
from functools import lru_cache
from transformers.models.llama.modeling_llama import logger
import os
torch_compile_options = {
@ -80,7 +78,8 @@ else:
# See https://github.com/pytorch-labs/attention-gym/blob/main/examples/flex_attn.ipynb
# for more examples
# BSD 3-Clause License Copyright (c) 2023, Driss Guessous, Horace He et al
import functools, math
import functools
import math
def generate_tanh_softcap(t):
def tanh_softcap(x, b, h, q_idx, kv_idx):

5
unsloth/kernels/fp8.py
View file

@ -13,11 +13,8 @@
# limitations under the License.
import os
import torch
import torch.nn as nn
import triton
import triton.language as tl
from torch.nn import functional as F
import math
from unsloth_zoo.utils import Version
from unsloth_zoo.log import logger
from unsloth_zoo.temporary_patches.common import torch_compile
@ -588,7 +585,7 @@ try:
_has_fbgemm = test_has_fbgemm()
if _has_fbgemm:
os.environ["UNSLOTH_HAS_FBGEMM"] = "1"
logger.info(f"Using fbgemm_gpu block quantized FP8 matmul")
logger.info("Using fbgemm_gpu block quantized FP8 matmul")
fp8_block_quant_linear = fp8_fbgemm_block_linear
else:
os.environ["UNSLOTH_HAS_FBGEMM"] = "0"

1
unsloth/kernels/geglu.py
View file

@ -16,7 +16,6 @@ import triton
import triton.language as tl
import torch
from .utils import (
calculate_settings,
triton_tanh,
torch_gpu_device,
)

3
unsloth/kernels/layernorm.py
View file

@ -17,9 +17,6 @@ import triton
import triton.language as tl
import torch
from .utils import calculate_settings, torch_gpu_device
from unsloth_zoo.patching_utils import (
patch_layernorm,
)
@triton.jit

3
unsloth/kernels/moe/autotune_cache.py
View file

@ -23,9 +23,8 @@ import json
import logging
import os
import time
from typing import Dict, List, Optional, Tuple, Any
from typing import Dict, Optional, Tuple, Any
import torch
import triton
logger = logging.getLogger(__name__)

2
unsloth/kernels/moe/benchmark/utils.py
View file

@ -178,7 +178,7 @@ def save_autotune_results(autotune_cache, mode, ref_time, fused_time, results_di
for key, config in autotune_cache.items():
key = [
str(k) if not "torch" in str(k) else str(k.split("torch.")[-1]) for k in key
str(k) if "torch" not in str(k) else str(k.split("torch.")[-1]) for k in key
]
filename = "_".join(key)
save_path = f"{save_dir}/{filename}.json"

1
unsloth/kernels/moe/grouped_gemm/interface.py
View file

@ -3,7 +3,6 @@
import logging
import warnings
from dataclasses import asdict
from unsloth import DEVICE_TYPE
import torch

1
unsloth/kernels/moe/grouped_gemm/kernels/autotuning.py
View file

@ -336,7 +336,6 @@ def exceeds_smem_capacity(
def common_prune_criteria(config: triton.Config, kwargs: dict, dtype):
from ..interface import supports_tma
from .tuning import get_device_properties
smem_size = get_device_properties().SIZE_SMEM

1
unsloth/kernels/moe/grouped_gemm/kernels/backward.py
View file

@ -1,7 +1,6 @@
# SPDX-License-Identifier: GNU Affero General Public License v3.0
# Copyright 2023-present the Unsloth team. All rights reserved.
import torch
import triton
import triton.language as tl

1
unsloth/kernels/moe/grouped_gemm/kernels/forward.py
View file

@ -1,7 +1,6 @@
# SPDX-License-Identifier: GNU Affero General Public License v3.0
# Copyright 2023-present the Unsloth team. All rights reserved.
import torch
import triton
import triton.language as tl

2
unsloth/kernels/moe/tests/test_grouped_gemm.py
View file

@ -7,7 +7,6 @@ import pytest
import torch
from grouped_gemm.interface import (
grouped_gemm,
grouped_gemm_dW,
grouped_gemm_dX,
grouped_gemm_forward,
@ -582,7 +581,6 @@ def _test_grouped_gemm_backward_dX(
kernel_config_bwd_dW = KernelConfigBackward_dW()
else:
from grouped_gemm.kernels.backward import (
_autotuned_grouped_gemm_dW_kernel,
_autotuned_grouped_gemm_dX_kernel,
)
from grouped_gemm.kernels.forward import (

1
unsloth/kernels/moe/tests/test_llama4_moe.py
View file

@ -2,7 +2,6 @@
# Copyright 2023-present the Unsloth team. All rights reserved.
import argparse
import sys
from contextlib import contextmanager
from functools import partial

2
unsloth/kernels/swiglu.py
View file

@ -15,7 +15,7 @@
import triton
import triton.language as tl
import torch
from .utils import calculate_settings, torch_gpu_device
from .utils import torch_gpu_device
# signed int32 max is 2**31-1 so num_elements cannot exceed 2**31
NUM_INT32_ELEMENTS = 2**31

5
unsloth/kernels/utils.py
View file

@ -19,18 +19,13 @@ import ctypes
MAX_FUSED_SIZE: int = 65536
next_power_of_2 = triton.next_power_of_2
import functools
from typing import Optional
from ..device_type import (
is_hip,
get_device_type,
DEVICE_TYPE,
DEVICE_TYPE_TORCH,
DEVICE_COUNT,
ALLOW_PREQUANTIZED_MODELS,
)
from .fp8 import weight_dequant, fp8_linear
import functools
# torch.cuda.amp.custom_fwd is deprecated >= 2.4
import torch

49
unsloth/models/_utils.py
View file

@ -86,35 +86,26 @@ from typing import Union, Optional, List, Any, Callable, Tuple, Iterator
from platform import system as platform_system
platform_system = platform_system()
import numpy as np
import contextlib
import re
from dataclasses import dataclass, field
import functools
import textwrap
import logging
import warnings, subprocess, inspect, psutil, os, math
import warnings
import inspect
import psutil
import os
from unsloth_zoo.utils import Version, get_quant_type
from importlib.metadata import version as importlib_version
from ..device_type import (
is_hip,
get_device_type,
DEVICE_TYPE,
DEVICE_TYPE_TORCH,
DEVICE_COUNT,
ALLOW_PREQUANTIZED_MODELS,
)
from ..import_fixes import UNSLOTH_ENABLE_LOGGING
from unsloth_zoo.log import logger
from unsloth_zoo.tokenizer_utils import (
patch_tokenizer as _patch_tokenizer,
)
from unsloth_zoo.rl_environments import (
check_python_modules,
create_locked_down_function,
execute_with_time_limit,
Benchmarker,
)
from unsloth_zoo.patching_utils import (
patch_compiling_bitsandbytes,
patch_layernorm,
@ -127,8 +118,6 @@ from unsloth_zoo.gradient_checkpointing import (
unsloth_offloaded_gradient_checkpoint,
patch_unsloth_gradient_checkpointing,
unpatch_unsloth_gradient_checkpointing,
Unsloth_Gradient_Checkpointer,
unsloth_gradient_checkpoint,
patch_gradient_checkpointing,
unpatch_gradient_checkpointing,
patch_unsloth_smart_gradient_checkpointing,
@ -313,7 +302,7 @@ class HideLoggingMessage(logging.Filter):
self.text = text
def filter(self, x):
return not (self.text in x.getMessage())
return self.text not in x.getMessage()
# Replace warning messages (analogous to HideLoggingMessage but for warnings.warn)
@ -737,21 +726,21 @@ def patch_mistral_nemo_config(config):
try:
# Some Config files use layer_type_validation
# for eg Gemma-2, so we must import it to stop errors.
from transformers.configuration_utils import layer_type_validation
pass
except:
pass
try:
# Transformers 5.0+ uses RotaryEmbeddingConfigMixin as a base class for configs
from transformers.modeling_rope_utils import RotaryEmbeddingConfigMixin
pass
except:
pass
from transformers import __version__ as transformers_version
try:
from transformers import PreTrainedConfig
pass
except:
from transformers import PretrainedConfig
pass
model_architectures = [
"llama",
@ -850,7 +839,7 @@ from transformers.utils import is_openai_available
if is_openai_available():
try:
from openai import OpenAI
pass
except:
print("Unsloth: OpenAI failed to import - ignoring for now.")
import transformers.utils
@ -862,9 +851,7 @@ if is_openai_available():
# =============================================
# Get Flash Attention v2 if Ampere (RTX 30xx, A100)
import bitsandbytes as bnb
from transformers import AutoTokenizer
from transformers.utils.import_utils import _is_package_available
SUPPORTS_BFLOAT16 = False
@ -882,9 +869,9 @@ if DEVICE_TYPE == "cuda":
try:
try:
# See https://github.com/unslothai/unsloth/issues/1437
from flash_attn.flash_attn_interface import flash_attn_gpu
pass
except:
from flash_attn.flash_attn_interface import flash_attn_cuda
pass
HAS_FLASH_ATTENTION = True
# Also check for softcapping
@ -931,9 +918,9 @@ elif DEVICE_TYPE == "hip":
try:
try:
# See https://github.com/unslothai/unsloth/issues/1437
from flash_attn.flash_attn_interface import flash_attn_gpu
pass
except:
from flash_attn.flash_attn_interface import flash_attn_cuda
pass
HAS_FLASH_ATTENTION = True
# Also check for softcapping
@ -1286,7 +1273,7 @@ USE_MODELSCOPE = os.environ.get("UNSLOTH_USE_MODELSCOPE", "0") == "1"
if USE_MODELSCOPE:
if importlib.util.find_spec("modelscope") is None:
raise ImportError(
f"You are using the modelscope hub, please install modelscope by `pip install modelscope -U`"
"You are using the modelscope hub, please install modelscope by `pip install modelscope -U`"
)
import socket
@ -1304,13 +1291,10 @@ def has_internet(host = "8.8.8.8", port = 53, timeout = 3):
return True
finally:
sock.close()
except socket.error as ex:
except socket.error:
return False
import psutil
def _get_statistics(statistics = None, force_download = True):
# We log some basic stats about which environment is being used.
# We simply download a README.md file from HF - all data is made public.
@ -1477,7 +1461,6 @@ def get_statistics(local_files_only = False):
# Fixes Bitsandbytes to remove missing warnings
from transformers.utils.quantization_config import (
BitsAndBytesConfig,
QuantizationMethod,
)
BitsAndBytesConfig__init__ = inspect.getsource(BitsAndBytesConfig.__init__)

4
unsloth/models/cohere.py
View file

@ -13,7 +13,6 @@
# limitations under the License.
from .llama import *
from ._utils import __version__
from unsloth_zoo.hf_utils import dtype_from_config
from unsloth_zoo.utils import _get_dtype, Version
from ..utils.packing import get_packed_info_from_kwargs
@ -30,9 +29,6 @@ try:
CohereDecoderLayer,
CohereModel,
CohereForCausalLM,
CohereRotaryEmbedding,
apply_rotary_pos_emb,
repeat_kv,
)
except:
transformers_version = Version(transformers_version)

4
unsloth/models/falcon_h1.py
View file

@ -13,8 +13,6 @@
# limitations under the License.
from .llama import *
import os
from ._utils import __version__
from unsloth_zoo.utils import Version, _get_dtype
from unsloth_zoo.hf_utils import dtype_from_config
from ..utils.packing import get_packed_info_from_kwargs
@ -28,7 +26,6 @@ from ..utils.attention_dispatch import (
from .llama import (
LlamaRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
_LlamaModel_fast_forward_inference,
)
try:
@ -37,7 +34,6 @@ try:
FalconH1DecoderLayer,
FalconH1Model,
FalconH1ForCausalLM,
FalconHybridMambaAttentionDynamicCache,
)
except:
from transformers import __version__ as transformers_version

9
unsloth/models/gemma.py
View file

@ -14,14 +14,8 @@
from .llama import *
from .llama import _get_rope_theta
from ._utils import __version__
from unsloth_zoo.utils import _get_dtype, Version
from unsloth_zoo.hf_utils import dtype_from_config
from ..utils.packing import (
build_sdpa_packed_attention_mask,
build_xformers_block_causal_mask,
get_packed_info_from_kwargs,
)
import math
try:
@ -30,9 +24,6 @@ try:
GemmaDecoderLayer,
GemmaModel,
GemmaForCausalLM,
GemmaRotaryEmbedding,
apply_rotary_pos_emb,
repeat_kv,
)
except:
transformers_version = Version(transformers_version)

7
unsloth/models/gemma2.py
View file

@ -13,7 +13,6 @@
# limitations under the License.
from .llama import *
from ._utils import __version__
from unsloth_zoo.utils import _get_dtype, Version
from unsloth_zoo.hf_utils import dtype_from_config
from ..utils.packing import get_packed_info_from_kwargs
@ -22,7 +21,6 @@ from ..utils.attention_dispatch import (
AttentionContext,
run_attention,
select_attention_backend,
SDPA,
)
from .gemma import (
GemmaFixedRotaryEmbedding,
@ -36,9 +34,6 @@ try:
Gemma2DecoderLayer,
Gemma2Model,
Gemma2ForCausalLM,
Gemma2RotaryEmbedding,
apply_rotary_pos_emb,
repeat_kv,
)
except:
transformers_version = Version(transformers_version)
@ -65,7 +60,7 @@ except:
Gemma2FlashAttention2 = Gemma2Attention
if HAS_FLASH_ATTENTION_SOFTCAPPING:
from flash_attn import flash_attn_func
pass
# Logit softcapping

9
unsloth/models/glm4_moe.py
View file

@ -25,20 +25,11 @@ Key architecture differences from Qwen3 MoE:
from .llama import *
import os
from ._utils import __version__
from .llama import (
LlamaRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
fix_prepare_inputs_for_generation,
fast_rms_layernorm_inference,
fast_swiglu_inference,
LlamaModel_fast_forward,
LlamaModel_fast_forward_inference,
CausalLM_fast_forward,
PeftModel_fast_forward,
)
import torch
import torch.nn.functional as F
from typing import Optional, Tuple
from ..kernels import fast_rms_layernorm

5
unsloth/models/granite.py
View file

@ -13,7 +13,6 @@
# limitations under the License.
from .llama import *
import os
from ._utils import __version__
from unsloth_zoo.utils import _get_dtype, Version
from unsloth_zoo.hf_utils import dtype_from_config
@ -265,8 +264,6 @@ def GraniteDecoderLayer_fast_forward(
return outputs
from math import sqrt as math_sqrt
KV_CACHE_INCREMENT = 256 # KV Cache update size
torch_nn_functional_softmax = torch.nn.functional.softmax
torch_matmul = torch.matmul
@ -285,7 +282,7 @@ def GraniteAttention_fast_forward_inference(
):
assert (
position_embeddings is not None
), f"Granite model requires position embeddings to be specified"
), "Granite model requires position embeddings to be specified"
Xn = hidden_states
bsz, _, hd = hidden_states.size()

15
unsloth/models/llama.py
View file

@ -26,7 +26,6 @@ from ._utils import (
_get_inference_mode_context_manager,
_prepare_model_for_qat,
is_bfloat16_supported,
get_quant_type,
)
from .loader_utils import _get_fp8_mode_and_check_settings
from ..utils.packing import (
@ -50,12 +49,9 @@ from unsloth_zoo.hf_utils import (
)
from unsloth_zoo.peft_utils import SKIP_QUANTIZATION_MODULES
from ..device_type import (
is_hip,
get_device_type,
DEVICE_TYPE,
DEVICE_TYPE_TORCH,
DEVICE_COUNT,
ALLOW_PREQUANTIZED_MODELS,
)
transformers_version = Version(transformers_version)
@ -97,7 +93,6 @@ except:
LlamaFlashAttention2 = LlamaAttention
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
AutoModelForSequenceClassification,
BitsAndBytesConfig,
@ -108,14 +103,16 @@ from transformers import set_seed as transformers_set_seed
from peft import LoraConfig, TaskType, get_peft_model as _get_peft_model
from peft import PeftModelForCausalLM, PeftModelForSequenceClassification
from ..save import patch_saving_functions
import re, os, inspect, math, sys
import re
import os
import inspect
import types
try:
from huggingface_hub.utils import get_token
pass
except:
# Old HF Hub versions <= 0.0.25
from huggingface_hub.utils._token import get_token
pass
from triton import __version__ as triton_version
HAS_XFORMERS = xformers is not None
@ -2983,7 +2980,7 @@ class FastLlamaModel:
try:
assert module in accepted_modules
final_modules.append(module)
except AssertionError as e:
except AssertionError:
final_modules.append(module)
print(
"Unsloth: You added custom modules, but Unsloth hasn't optimized for this.\n"

21
unsloth/models/loader.py
View file

@ -15,20 +15,17 @@
from ._utils import (
_prepare_model_for_qat,
is_bfloat16_supported,
is_vLLM_available,
HAS_FLASH_ATTENTION,
HAS_FLASH_ATTENTION_SOFTCAPPING,
USE_MODELSCOPE,
get_transformers_model_type,
hf_login,
)
from .granite import FastGraniteModel
from .llama import FastLlamaModel, logger
from .mistral import FastMistralModel
from .qwen2 import FastQwen2Model
from .qwen3 import FastQwen3Model
from .qwen3_moe import FastQwen3MoeModel
from .cohere import FastCohereModel
from transformers import AutoConfig
from transformers import __version__ as transformers_version
from peft import PeftConfig, PeftModel
@ -39,22 +36,20 @@ from .loader_utils import (
get_model_name,
prepare_device_map,
)
import os, contextlib, sys
import os
import contextlib
try:
from huggingface_hub import get_token
pass
except:
try:
from huggingface_hub.utils import get_token
pass
except:
# For older versions of huggingface_hub
from huggingface_hub.utils._token import get_token
pass
from huggingface_hub import HfFileSystem
import importlib.util
from ..device_type import (
is_hip,
get_device_type,
DEVICE_TYPE,
DEVICE_TYPE_TORCH,
DEVICE_COUNT,
ALLOW_PREQUANTIZED_MODELS,
@ -86,15 +81,12 @@ if SUPPORTS_GEMMA:
if SUPPORTS_GEMMA2:
from .gemma2 import FastGemma2Model
if SUPPORTS_FALCON_H1:
from .falcon_h1 import FastFalconH1Model
pass
import torch
from ._utils import (
patch_compiling_bitsandbytes,
patch_model_and_tokenizer,
prepare_model_for_kbit_training,
apply_unsloth_gradient_checkpointing,
patch_compiled_autograd,
process_vision_info,
unsloth_compile_transformers,
fast_inference_setup,
)
@ -806,7 +798,6 @@ class FastLanguageModel(FastLlamaModel):
from ..kernels import (
patch_loss_functions,
post_patch_loss_function,
)
from .vision import FastBaseModel
from transformers import (

1
unsloth/models/loader_utils.py
View file

@ -16,7 +16,6 @@ from ..device_type import DEVICE_TYPE_TORCH
import importlib
import os
import torch
import re
import tempfile
from typing import Union
from .mapper import (

644
unsloth/models/mapper.py
View file

File diff suppressed because it is too large Load diff

2
unsloth/models/mistral.py
View file

@ -14,7 +14,6 @@
from .llama import *
import os
from ._utils import __version__
from unsloth_zoo.utils import _get_dtype
from unsloth_zoo.hf_utils import dtype_from_config
from ..utils.packing import (
@ -48,7 +47,6 @@ try:
except:
MistralSdpaAttention = MistralAttention
MistralFlashAttention2 = MistralAttention
from unsloth_zoo.utils import Version, _get_dtype
def MistralAttention_fast_forward(

7
unsloth/models/qwen3.py
View file

@ -13,9 +13,7 @@
# limitations under the License.
from .llama import *
import os
from ._utils import __version__
from unsloth_zoo.utils import Version, _get_dtype
from unsloth_zoo.utils import Version
from ..utils.packing import get_packed_info_from_kwargs
from ..utils.attention_dispatch import (
AttentionConfig,
@ -48,9 +46,6 @@ except:
f'Try `pip install --upgrade "transformers>=4.50.3"`\n'
f"to obtain the latest transformers build, then restart this session."
)
from transformers.modeling_attn_mask_utils import (
_prepare_4d_causal_attention_mask_for_sdpa,
)
# For Pytorch 2.1.1
try:

3
unsloth/models/qwen3_moe.py
View file

@ -13,8 +13,6 @@
# limitations under the License.
from .llama import *
import os
from ._utils import __version__
from .llama import (
LlamaRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
@ -43,7 +41,6 @@ from transformers.models.qwen3_moe.modeling_qwen3_moe import (
# Qwen3SdpaAttention = Qwen3Attention
# Qwen3FlashAttention2 = Qwen3Attention
# pass
from unsloth_zoo.utils import Version, _get_dtype
torch_nn_functional_softmax = torch.nn.functional.softmax

3
unsloth/models/rl.py
View file

@ -18,7 +18,6 @@ __all__ = [
]
import torch
from typing import Any, Callable, Dict, List, Literal, Optional, Tuple, Union
import inspect
import os
import re
@ -539,8 +538,6 @@ def _wrap_grpo_generate_and_score(trainer_cls):
def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
# Patch for vLLM and Unsloth PEFT
import trl
import trl.trainer
try:
trainer = eval(f"trl.trainer.{trainer_file}")

5
unsloth/models/rl_replacements.py
View file

@ -38,12 +38,7 @@ from unsloth_zoo.log import logger
from unsloth_zoo.device_type import device_synchronize
import importlib.util
from ..device_type import (
is_hip,
get_device_type,
DEVICE_TYPE,
DEVICE_TYPE_TORCH,
DEVICE_COUNT,
ALLOW_PREQUANTIZED_MODELS,
)
import textwrap
from ._utils import _get_inference_mode_context_manager

6
unsloth/models/sentence_transformer.py
View file

@ -384,7 +384,7 @@ def _push_to_hub_gguf(
# Save to temporary directory first
with tempfile.TemporaryDirectory(prefix = "unsloth_st_gguf_") as temp_dir:
print(f"Unsloth: Converting SentenceTransformer to GGUF format...")
print("Unsloth: Converting SentenceTransformer to GGUF format...")
# Call save_pretrained_gguf to do the local conversion
result = _save_pretrained_gguf(
@ -577,9 +577,9 @@ class FastSentenceTransformer(FastModel):
print(f"Pooling mode detected as {mode}, updating...")
return mode
except Exception as e:
except Exception:
print(
f"Failed to detect pooling mode, not a sentence-transformers model. Using default pooling mode 'mean', this may or may not work."
"Failed to detect pooling mode, not a sentence-transformers model. Using default pooling mode 'mean', this may or may not work."
)
return "mean"

Some files were not shown because too many files have changed in this diff Show more