Merge branch 'master' into issue-20801-cli-subcommand-ckpt-path-hparams

Author: Borda

docs/source-pytorch/versioning.rst

@@ -53,12 +53,8 @@ API Evolution
 
 Lightning's development is driven by research and best practices in a rapidly developing field of AI and machine learning. Change is inevitable and when it happens, the Lightning team is committed to minimizing user friction and maximizing ease of transition from one version to the next. We take backwards compatibility and reproducibility very seriously.
 
-For API removal, renaming or other forms of backwards-incompatible changes, the procedure is:
-
-#. A deprecation process is initiated at a minor version ``MAJOR.MINOR.PATCH`` (e.g. ``1.5.0``), producing a deprecation warning at runtime and removing it from the documentation.
-#. The deprecated API remains unchanged during the deprecation phase for two minor versions or the next major update, whichever comes first.
-#. The breaking change is done in version ``MAJOR.(MINOR+2).0`` (e.g. ``1.7.0``), or ``(MAJOR+1).0.0`` (e.g. ``2.0.0``), whichever comes first.
-#. From that version onward, the deprecation warning gets converted into a helpful error, which will remain until next major release.
+Excepting extenuating circumstances (e.g. a critical bug), API removal, renaming or other forms of backwards-incompatible changes are limited to major version upgrades — that is ``(MAJOR+1).0.0``.
+Concretely, a breaking change for an API introduced in ``2.x.x`` can be introduced with Lightning ``3.0.0``.
 
 This policy is not strict. Shorter or longer deprecation cycles may apply to some cases.
 For example, in the past DDP2 was removed without a deprecation process because the feature was broken and unusable beyond fixing as discussed in `#12584 <https://github.com/Lightning-AI/pytorch-lightning/issues/12584>`_.
@@ -69,6 +65,14 @@ Compatibility matrix
 
 PyTorch Lightning follows `NEP 29 <https://numpy.org/neps/nep-0029-deprecation_policy.html>`_ which PyTorch also follows (`#74203 <https://github.com/pytorch/pytorch/issues/74203>`_).
 The table below indicates the coverage of tested versions in our CI. Versions outside the ranges may unofficially work in some cases.
+Since the release of PyTorch `2.0`, Lightning strives to officially support the latest 5 PyTorch minor releases with no breaking changes within major versions [1]_.
+
+
+.. note::
+   Legend used in the table below:
+
+   - "≥ X.Y" minimum supported version.
+   - "▼ X.Y" last CI-tested version (informational, not a hard upper bound).
 
 .. list-table::
    :header-rows: 1
@@ -82,102 +86,104 @@ The table below indicates the coverage of tested versions in our CI. Versions ou
    * - 2.5
      - 2.5
      - 2.5
-     - ≥2.1, ≤2.7
+     - ≥2.1 (▼ 2.8)
      - ≥0.7.0
-     - ≥3.9, ≤3.12
+     - ≥3.9 (▼ 3.12)
    * - 2.4
      - 2.4
      - 2.4
-     - ≥2.1, ≤2.6
+     - ≥2.1 (▼ 2.6)
      - ≥0.7.0
-     - ≥3.9, ≤3.12
+     - ≥3.9 (▼ 3.12)
    * - 2.3
      - 2.3
      - 2.3
-     - ≥2.0, ≤2.3
+     - ≥2.0 (▼ 2.3)
      - ≥0.7.0
-     - ≥3.8, ≤3.11
+     - ≥3.8 (▼ 3.11)
    * - 2.2
      - 2.2
      - 2.2
-     - ≥1.13, ≤2.2
+     - ≥1.13 (▼ 2.2)
      - ≥0.7.0
-     - ≥3.8, ≤3.11
+     - ≥3.8 (▼ 3.11)
    * - 2.1
      - 2.1
      - 2.1
-     - ≥1.12, ≤2.1
+     - ≥1.12 (▼ 2.1)
      - ≥0.7.0
-     - ≥3.8, ≤3.11
+     - ≥3.8 (▼ 3.11)
    * - 2.0
      - 2.0
      - 2.0 (GA)
-     - ≥1.11, ≤2.0
+     - ≥1.11 (▼ 2.0)
      - ≥0.7.0
-     - ≥3.8, ≤3.10
+     - ≥3.8 (▼ 3.10)
    * - 1.9
      - 1.9
      - 1.9 (experimental)
-     - ≥1.10, ≤1.13
+     - ≥1.10 (▼ 1.13)
      - ≥0.7.0
-     - ≥3.7, ≤3.10
+     - ≥3.7 (▼ 3.10)
    * - 1.8**
      - 1.8
      - n/a***
-     - ≥1.10, ≤1.13
+     - ≥1.10 (▼ 1.13)
      - ≥0.7.0
-     - ≥3.7, ≤3.10
+     - ≥3.7 (▼ 3.10)
    * - n/a
      - 1.7
      - n/a***
-     - ≥1.9, ≤1.12
+     - ≥1.9 (▼ 1.12)
      - ≥0.7.0
-     - ≥3.7, ≤3.10
+     - ≥3.7 (▼ 3.10)
    * - n/a
      - 1.6
      - n/a***
-     - ≥1.8, ≤1.11
+     - ≥1.8 (▼ 1.11)
      - ≥0.4.1
-     - ≥3.7, ≤3.9
+     - ≥3.7 (▼ 3.9)
    * - n/a
      - 1.5
      - n/a***
-     - ≥1.7, ≤1.10
+     - ≥1.7 (▼ 1.10)
      - ≥0.4.1
-     - ≥3.6, ≤3.9
+     - ≥3.6 (▼ 3.9)
    * - n/a
      - 1.4
      - n/a
-     - ≥1.6, ≤1.9
+     - ≥1.6 (▼ 1.9)
      - ≥0.4.0
-     - ≥3.6, ≤3.9
+     - ≥3.6 (▼ 3.9)
    * - n/a
      - 1.3
      - n/a
-     - ≥1.4, ≤1.8
+     - ≥1.4 (▼ 1.8)
      - ≥0.2.0
-     - ≥3.6, ≤3.9
+     - ≥3.6 (▼ 3.9)
    * - n/a
      - 1.2
      - n/a
-     - ≥1.4, ≤1.8
+     - ≥1.4 (▼ 1.8)
      - n/a*
-     - ≥3.6, ≤3.8
+     - ≥3.6 (▼ 3.8)
    * - n/a
      - 1.1
      - n/a
-     - ≥1.3, ≤1.8
+     - ≥1.3 (▼ 1.8)
      - n/a*
-     - ≥3.6, ≤3.8
+     - ≥3.6 (▼ 3.8)
    * - n/a
      - 1.0
      - n/a
-     - ≥1.3, ≤1.7
+     - ≥1.3 (▼ 1.7)
      - n/a*
-     - ≥3.6, ≤3.8
+     - ≥3.6 (▼ 3.8)
 
 \* ``torchmetrics`` was part of ``pytorch_lightning`` at the time and was decoupled to a separate package in v1.3.
 
 \*\* The joint ``lightning`` package was first published in version 1.8
 
 \*\*\* Fabric is the evolution of ``LightningLite`` which was released inside ``pytorch_lightning`` 1.5 and was decoupled to a separate package in v1.9
+
+.. [1] See `this community discussion <https://github.com/Lightning-AI/pytorch-lightning/issues/21073#issuecomment-3201706857>`_.

requirements/pytorch/test.txt

@@ -19,5 +19,4 @@ uvicorn  # for `ServableModuleValidator`  # not setting version as re-defined in
 
 tensorboard >=2.11, <2.21.0  # for `TensorBoardLogger`
 
---find-links https://download.pytorch.org/whl/torch-tensorrt
 torch-tensorrt; platform_system == "Linux" and python_version >= "3.12"

src/lightning/fabric/CHANGELOG.md

@@ -29,6 +29,14 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ---
 
+## [2.5.4] - 2025-08-29
+
+### Changed
+
+- Added support for NVIDIA H200 GPUs in `get_available_flops` ([#20913](https://github.com/Lightning-AI/pytorch-lightning/pull/21119))
+
+
+
 ## [2.5.3] - 2025-08-13
 
 ### Changed

src/lightning/fabric/utilities/throughput.py

@@ -304,6 +304,23 @@ def measure_flops(
 
 _CUDA_FLOPS: dict[str, dict[Union[str, torch.dtype], float]] = {
     # Hopper
+    # source: https://nvdam.widen.net/s/nb5zzzsjdf/hpc-datasheet-sc23-h200-datasheet-3002446
+    "h200 sxm1": {
+        torch.float64: 3.4e13,
+        torch.float32: 6.7e13,
+        "tfloat32": 9.9e14,
+        torch.bfloat16: 2.0e15,
+        torch.float16: 2.0e15,
+        torch.int8: 4.0e15,
+    },
+    "h200 nvl1": {
+        torch.float64: 3.0e13,
+        torch.float32: 6.0e13,
+        "tfloat32": 8.4e14,
+        torch.bfloat16: 1.7e15,
+        torch.float16: 1.7e15,
+        torch.int8: 3.3e15,
+    },
     # source: https://resources.nvidia.com/en-us-tensor-core
     "h100 nvl": {
         torch.float64: 67e12,
@@ -536,7 +553,12 @@ def get_available_flops(device: torch.device, dtype: Union[torch.dtype, str]) ->
     if device.type == "cuda":
         device_name = torch.cuda.get_device_name(device)
         chip = device_name.lower()
-        if "h100" in chip:
+        if "h200" in chip:
+            if "sxm1" in chip:
+                chip = "h200 sxm1"
+            elif "nvl1" in chip:
+                chip = "h200 nvl1"
+        elif "h100" in chip:
             if "hbm3" in chip:
                 chip = "h100 sxm"
             elif "nvl" in chip:

src/lightning/pytorch/CHANGELOG.md

@@ -18,7 +18,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
-- Default to RichProgressBar and RichModelSummary if the rich package is available. Fallback to TQDMProgressBar and ModelSummary otherwise. ([#9580](https://github.com/Lightning-AI/pytorch-lightning/pull/9580))
+- Default to `RichProgressBar` and `RichModelSummary` if the rich package is available. Fallback to TQDMProgressBar and ModelSummary otherwise. ([#9580](https://github.com/Lightning-AI/pytorch-lightning/pull/9580))
 
 
 ### Removed
@@ -28,24 +28,26 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
-- Fixed `AsyncCheckpointIO` snapshots tensors to avoid race with parameter mutation ([#21079](https://github.com/Lightning-AI/pytorch-lightning/pull/21079))
+- Fixed `LightningCLI` not using `ckpt_path` hyperparameters to instantiate classes ([#21116](https://github.com/Lightning-AI/pytorch-lightning/pull/21116))
 
 
-- Fixed `AsyncCheckpointIO` threadpool exception if calling fit or validate more than one ([#20952](https://github.com/Lightning-AI/pytorch-lightning/pull/20952))
+-
 
 
-- Fixed learning rate not being correctly set after using `LearningRateFinder` callback ([#21068](https://github.com/Lightning-AI/pytorch-lightning/pull/21068))
+---
 
 
-- Fixed misalignment column while using rich model summary in `DeepSpeedstrategy` ([#21100](https://github.com/Lightning-AI/pytorch-lightning/pull/21100))
+## [2.5.4] - 2025-08-29
 
+### Fixed
 
+- Fixed `AsyncCheckpointIO` snapshots tensors to avoid race with parameter mutation ([#21079](https://github.com/Lightning-AI/pytorch-lightning/pull/21079))
+- Fixed `AsyncCheckpointIO` threadpool exception if calling fit or validate more than one ([#20952](https://github.com/Lightning-AI/pytorch-lightning/pull/20952))
+- Fixed learning rate not being correctly set after using `LearningRateFinder` callback ([#21068](https://github.com/Lightning-AI/pytorch-lightning/pull/21068))
+- Fixed misalignment column while using rich model summary in `DeepSpeedstrategy` ([#21100](https://github.com/Lightning-AI/pytorch-lightning/pull/21100))
 - Fixed `RichProgressBar` crashing when sanity checking using val dataloader with 0 len ([#21108](https://github.com/Lightning-AI/pytorch-lightning/pull/21108))
 
 
-- Fixed `LightningCLI` not using `ckpt_path` hyperparameters to instantiate classes ([#21116](https://github.com/Lightning-AI/pytorch-lightning/pull/21116))
-
-
 ---
 
 ## [2.5.3] - 2025-08-13

tests/legacy/back-compatible-versions.txt

@@ -107,3 +107,4 @@
 2.5.1
 2.5.2
 2.5.3
+2.5.4

tests/tests_fabric/utilities/test_throughput.py

@@ -68,6 +68,8 @@ def test_get_available_flops(xla_available):
     "device_name",
     [
         # Hopper
+        "NVIDIA H200 SXM1",
+        "NVIDIA H200 NVL1",
         "h100-nvl",  # TODO: switch with `torch.cuda.get_device_name()` result
         "h100-hbm3",  # TODO: switch with `torch.cuda.get_device_name()` result
         "NVIDIA H100 PCIe",