bump PT 2.8 & install with --upgrade-strategy greedy (#2107)

Borda · pre-commit-ci[bot] · web-flow · commit f464693df3fd · 2025-09-09T08:09:54.000+02:00
Co-authored-by: pre-commit-ci[bot] &lt;66853113+pre-commit-ci[bot]@users.noreply.github.com&gt;
diff --git a/.azure/gpu-test.yml b/.azure/gpu-test.yml
@@ -29,10 +29,13 @@ jobs:
       HF_HOME: "/var/tmp/hf/home"
       HF_HUB_CACHE: "/var/tmp/hf/hub"
       SKIP_WITH_CI: "1"
+      SKIP_WITH_GPT_CI: "1"
       NCCL_DEBUG: "INFO"
+      TORCHDYNAMO_VERBOSE: "1"
+      TORCHDYNAMO_INLINE_INBUILT_NN_MODULES: "0"
       PYTHON_VERSION: "3.10"
       CUDA_VERSION: "12.6.3"
-      TORCH_VERSION: "2.7.1"
+      TORCH_VERSION: "2.8.0"
       CUDNN_FRONTEND_VERSION: "1.10.0"
       CUBLAS_WORKSPACE_CONFIG: ":4096:8"
     container:
@@ -64,7 +67,7 @@ jobs:
 
       - script: |
           pip install --upgrade pip
-          pip install '.[extra,test]' "torch==${TORCH_VERSION}" cffi -U
+          pip install '.[extra,test]' "torch==${TORCH_VERSION}" cffi -U --upgrade-strategy eager
         displayName: "Install package & dependencies"
 
       - script: |
@@ -95,7 +98,6 @@ jobs:
           PL_RUN_STANDALONE_TESTS: "1"
           # NUM_PARALLEL_TESTS: "10"
           NCCL_IGNORE_DISABLED_P2P: "1"
-          NCCL_DEBUG: "INFO"
         timeoutInMinutes: "10"
 
       - bash: |
@@ -110,6 +112,4 @@ jobs:
           RUN_ONLY_CUDA_TESTS=0 pytest tests/ext_thunder/test_thunder_networks.py -v --durations=50
         displayName: "Extra tests for Thunder [main branch]"
         condition: eq(variables['dependency'], 'compiler')
-        env:
-          TORCHDYNAMO_VERBOSE: "1"
         timeoutInMinutes: "10"
diff --git a/.github/workflows/cpu-tests.yml b/.github/workflows/cpu-tests.yml
@@ -119,7 +119,7 @@ jobs:
           python -m lightning_utilities.cli requirements set-oldest --req_files=pyproject.toml
       - name: Install dependencies
         run: |
-          pip install '.[extra,compiler,test]' -U
+          pip install '.[extra,compiler,test]' -U --upgrade-strategy eager
           pip list
 
       - name: Run tests
diff --git a/litgpt/data/alpaca.py b/litgpt/data/alpaca.py
@@ -7,12 +7,12 @@
 from typing import Optional, Union
 
 import torch
-from lightning_utilities.core.imports import RequirementCache
 from torch.utils.data import DataLoader, random_split
 
 from litgpt.data.base import DataModule, SFTDataset, get_sft_collate_fn
 from litgpt.prompts import PromptStyle
 from litgpt.tokenizer import Tokenizer
+from litgpt.utils import _REQUESTS_AVAILABLE
 
 _URL = "https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_cleaned_archive.json"
 
@@ -115,9 +115,8 @@ def download_if_missing(file_path: Path, file_url: str, mode: str = "w", stream:
     """Downloads the raw json data file and saves it in the given destination."""
     if file_path.exists() and file_path.stat().st_size > 0:
         return
-    requests_available = RequirementCache("requests")
-    if not requests_available:
-        raise ModuleNotFoundError(str(requests_available))
+    if not _REQUESTS_AVAILABLE:
+        raise ModuleNotFoundError(str(_REQUESTS_AVAILABLE))
     import requests
 
     response = requests.get(file_url, stream=stream)
diff --git a/litgpt/data/prepare_starcoder.py b/litgpt/data/prepare_starcoder.py
@@ -5,12 +5,9 @@
 import traceback
 from pathlib import Path
 
-from lightning_utilities.core.imports import RequirementCache
-
 from litgpt.tokenizer import Tokenizer
-from litgpt.utils import CLI, extend_checkpoint_dir
+from litgpt.utils import _LITDATA_AVAILABLE, CLI, extend_checkpoint_dir
 
-_LITDATA_AVAILABLE = RequirementCache("litdata")
 if _LITDATA_AVAILABLE:
     from litdata.processing.data_processor import DataChunkRecipe
 else:
diff --git a/litgpt/deploy/serve.py b/litgpt/deploy/serve.py
@@ -6,13 +6,10 @@
 from typing import Any, Dict, Literal, Optional
 
 import torch
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt.api import LLM
-from litgpt.utils import auto_download_checkpoint
+from litgpt.utils import _JINJA2_AVAILABLE, _LITSERVE_AVAILABLE, auto_download_checkpoint
 
-_LITSERVE_AVAILABLE = RequirementCache("litserve")
-_JINJA2_AVAILABLE = RequirementCache("jinja2")
 if _LITSERVE_AVAILABLE:
     from litserve import LitAPI, LitServer
     from litserve.specs.openai import ChatCompletionRequest, OpenAISpec
diff --git a/litgpt/finetune/adapter.py b/litgpt/finetune/adapter.py
@@ -13,7 +13,6 @@
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.strategies import FSDPStrategy
 from lightning.fabric.utilities import ThroughputMonitor
-from lightning_utilities.core.imports import RequirementCache
 from torch.utils.data import ConcatDataset, DataLoader
 from torchmetrics import RunningMean
 
@@ -24,6 +23,7 @@
 from litgpt.prompts import save_prompt_style
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     CycleIterator,
     auto_download_checkpoint,
     check_nvlink_connectivity,
@@ -108,7 +108,7 @@ def setup(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/finetune/adapter_v2.py b/litgpt/finetune/adapter_v2.py
@@ -13,7 +13,6 @@
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.strategies import FSDPStrategy
 from lightning.fabric.utilities import ThroughputMonitor
-from lightning_utilities.core.imports import RequirementCache
 from torch.utils.data import ConcatDataset, DataLoader
 from torchmetrics import RunningMean
 
@@ -24,6 +23,7 @@
 from litgpt.prompts import save_prompt_style
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     CycleIterator,
     auto_download_checkpoint,
     check_nvlink_connectivity,
@@ -110,7 +110,7 @@ def setup(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/finetune/lora.py b/litgpt/finetune/lora.py
@@ -13,7 +13,6 @@
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.strategies import ModelParallelStrategy
 from lightning.fabric.utilities import ThroughputMonitor
-from lightning_utilities.core.imports import RequirementCache
 from torch.utils.data import ConcatDataset, DataLoader
 from torchmetrics import RunningMean
 
@@ -25,6 +24,7 @@
 from litgpt.scripts.merge_lora import merge_lora
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     CycleIterator,
     auto_download_checkpoint,
     check_nvlink_connectivity,
@@ -140,7 +140,7 @@ def setup(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/finetune/lora_legacy.py b/litgpt/finetune/lora_legacy.py
@@ -13,7 +13,6 @@
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.strategies import FSDPStrategy
 from lightning.fabric.utilities import ThroughputMonitor
-from lightning_utilities.core.imports import RequirementCache
 from torch.utils.data import ConcatDataset, DataLoader
 from torchmetrics import RunningMean
 
@@ -25,6 +24,7 @@
 from litgpt.scripts.merge_lora import merge_lora
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     CycleIterator,
     auto_download_checkpoint,
     check_nvlink_connectivity,
@@ -139,7 +139,7 @@ def setup(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/adapter.py b/litgpt/generate/adapter.py
@@ -10,13 +10,13 @@
 import lightning as L
 import torch
 from lightning.fabric.plugins import BitsandbytesPrecision
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt import PromptStyle, Tokenizer
 from litgpt.adapter import GPT, Config
 from litgpt.generate.base import generate
 from litgpt.prompts import has_prompt_style, load_prompt_style
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_file_size_on_cpu_and_warn,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -83,7 +83,7 @@ def main(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/adapter_v2.py b/litgpt/generate/adapter_v2.py
@@ -10,13 +10,13 @@
 import lightning as L
 import torch
 from lightning.fabric.plugins import BitsandbytesPrecision
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt import PromptStyle, Tokenizer
 from litgpt.adapter_v2 import GPT, Config
 from litgpt.generate.base import generate
 from litgpt.prompts import has_prompt_style, load_prompt_style
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_file_size_on_cpu_and_warn,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -83,7 +83,7 @@ def main(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/base.py b/litgpt/generate/base.py
@@ -12,13 +12,13 @@
 import torch._dynamo.config
 import torch._inductor.config
 from lightning.fabric.plugins import BitsandbytesPrecision
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt.config import Config
 from litgpt.model import GPT
 from litgpt.prompts import PromptStyle, has_prompt_style, load_prompt_style
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_file_size_on_cpu_and_warn,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -485,7 +485,7 @@ def main(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/full.py b/litgpt/generate/full.py
@@ -10,12 +10,12 @@
 import lightning as L
 import torch
 from lightning.fabric.plugins import BitsandbytesPrecision
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt import GPT, Config, PromptStyle, Tokenizer
 from litgpt.generate.base import generate
 from litgpt.prompts import has_prompt_style, load_prompt_style
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_file_size_on_cpu_and_warn,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -82,7 +82,7 @@ def main(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/sequentially.py b/litgpt/generate/sequentially.py
@@ -18,15 +18,19 @@
 from lightning.fabric.accelerators import CUDAAccelerator
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.utilities.init import _materialize_meta_tensors
-from lightning_utilities.core.imports import RequirementCache
 from tqdm import tqdm
 
 import litgpt.generate.base as generate_base
 from litgpt.config import Config
 from litgpt.model import GPT, Block, build_mask_cache
 from litgpt.prompts import PromptStyle, has_prompt_style, load_prompt_style
 from litgpt.tokenizer import Tokenizer
-from litgpt.utils import check_valid_checkpoint_dir, extend_checkpoint_dir, get_default_supported_precision
+from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
+    check_valid_checkpoint_dir,
+    extend_checkpoint_dir,
+    get_default_supported_precision,
+)
 
 
 @torch.inference_mode()
@@ -192,7 +196,7 @@ def main(
             raise NotImplementedError  # untested
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/speculative_decoding.py b/litgpt/generate/speculative_decoding.py
@@ -13,14 +13,14 @@
 import torch._inductor.config
 import torch.nn.functional as F
 from lightning.fabric.plugins import BitsandbytesPrecision
-from lightning_utilities.core.imports import RequirementCache
 
 from litgpt.config import Config
 from litgpt.generate.base import multinomial_num_samples_1, next_token, sample_top_p
 from litgpt.model import GPT
 from litgpt.prompts import PromptStyle, has_prompt_style, load_prompt_style
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_file_size_on_cpu_and_warn,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -388,7 +388,7 @@ def main(
     if quantize is not None and quantize.startswith("bnb."):
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/generate/tp.py b/litgpt/generate/tp.py
@@ -15,14 +15,14 @@
 import torch._inductor.config
 from lightning.fabric.plugins import BitsandbytesPrecision
 from lightning.fabric.utilities import rank_zero_only
-from lightning_utilities.core.imports import RequirementCache
 
 import litgpt.generate.base as generate_base
 from litgpt.config import Config
 from litgpt.model import GPT, CausalSelfAttention, GptNeoxMLP, LLaMAMLP, LLaMAMoE
 from litgpt.prompts import PromptStyle, has_prompt_style, load_prompt_style
 from litgpt.tokenizer import Tokenizer
 from litgpt.utils import (
+    _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0,
     check_nvlink_connectivity,
     check_valid_checkpoint_dir,
     extend_checkpoint_dir,
@@ -158,7 +158,7 @@ def main(
             raise NotImplementedError  # untested
         if "mixed" in precision:
             raise ValueError("Quantization and mixed precision is not supported.")
-        if RequirementCache("bitsandbytes != 0.42.0"):
+        if _BITANDBYTES_AVAILABLE_NOT_EQUAL_0_42_0:
             warnings.warn(
                 "LitGPT only supports bitsandbytes v0.42.0. This may result in errors when using quantization."
             )
diff --git a/litgpt/pretrain.py b/litgpt/pretrain.py
@@ -24,6 +24,8 @@
 from litgpt.data import DataModule, TinyLlama
 from litgpt.model import GPT, Block, CausalSelfAttention, Config, LLaMAMLP
 from litgpt.utils import (
+    _TORCH_EQUAL_2_7,
+    _TORCH_EQUAL_2_8,
     CycleIterator,
     capture_hparams,
     check_nvlink_connectivity,
@@ -238,9 +240,9 @@ def main(
 
     # work around PyTorch issue https://github.com/pytorch/pytorch/issues/152162
     # which does not like the lazy initialization to be called in dynamo.
-    # Happens with PyTorch 2.7.
+    # TODO: Happens with PyTorch 2.7+
     if (
-        torch.__version__.startswith("2.7.")
+        (_TORCH_EQUAL_2_7 or _TORCH_EQUAL_2_8)
         and (model._forward_module.__class__.__name__ == "OptimizedModule")
         and (model._forward_module._orig_mod.__class__.__name__ == "FullyShardedDataParallel")
     ):
diff --git a/litgpt/scripts/convert_hf_checkpoint.py b/litgpt/scripts/convert_hf_checkpoint.py
diff --git a/litgpt/scripts/download.py b/litgpt/scripts/download.py
diff --git a/litgpt/utils.py b/litgpt/utils.py
