[PyTorch] `GroupedTensor` integration #2600

ksivaman · 2026-01-15T14:58:31Z

Description

#2388 introduced the GroupedTensor class in the core library. This PR partly integrates this functionality to the PyTorch bindings.

Type of change

Documentation change (change only to the documentation, either a fix or a new content)
Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Breaking change (fix or feature that would cause existing functionality to not work as expected)
Infra/Build change
Code refactoring

Changes

Expose a python GroupedTensor class.
Integrate GroupedTensor into GroupedLinear such that the parameters are contiguous.
Expose a C++ grouped_quantize API to python similar to the split_quantize which returns a quantized GroupedTensor that can be directly consumed by the GEMMs ([common] Add support for cuBLASLt GEMM for GroupedTensor #2502).

Checklist:

I have read and followed the contributing guidelines
The functionality is complete
I have commented my code, particularly in hard-to-understand areas
I have made corresponding changes to the documentation
My changes generate no new warnings
I have added tests that prove my fix is effective or that my feature works
New and existing unit tests pass locally with my changes

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

greptile-apps · 2026-01-15T15:02:34Z

Greptile Summary

This PR integrates the GroupedTensor class from #2388 into PyTorch bindings, enabling contiguous memory storage for multiple weight tensors in GroupedLinear.

Key Changes

New GroupedTensor class (918 lines): Stores multiple tensors with different shapes in contiguous memory, supporting all quantization recipes (FP8, MXFP8, NVFP4, block scaling)
GroupedLinear integration: Added make_grouped_weights() method that converts individual weight parameters into views of a single contiguous GroupedTensor storage
Recipe API refactoring: Changed type-checking methods from instance methods to classmethods (isinstance → issubclass) to align with _get_compatible_recipe() returning class types
Quantizer enhancements: Added get_columnwise_shape() and get_scale_shape() methods for proper memory layout calculations
Comprehensive tests: 430-line test suite verifying contiguous memory layout and quantization correctness across all recipes

Implementation Notes

The implementation allocates all weight data in a single contiguous buffer, then creates individual parameter views that share the underlying storage. This improves memory locality and enables future optimizations like grouped GEMMs (#2502).

Confidence Score: 4/5

This PR is safe to merge with minor caveats that should be verified through testing
The implementation is well-designed and comprehensive, with extensive tests covering all quantization recipes. The core GroupedTensor logic is sound, and the integration into GroupedLinear follows established patterns. However, there are two acknowledged areas needing verification: (1) the copy operation from regular tensors to quantized tensors in make_grouped_weights() has a TODO comment indicating uncertainty about correctness across all recipes, and (2) the assumption that all quantizers in a group are "effectively the same" is not strongly enforced. The recipe API change from instance methods to classmethods is correct but represents a subtle behavioral change that could affect code calling these methods on instances.
Pay close attention to transformer_engine/pytorch/module/grouped_linear.py (copy operation in make_grouped_weights) and transformer_engine/common/recipe/__init__.py (instance method to classmethod change)

Important Files Changed

Filename	Overview
transformer_engine/pytorch/tensor/storage/grouped_tensor.py	New 918-line file implementing GroupedTensor class for contiguous storage of multiple tensors with different shapes. Comprehensive implementation with quantization support for FP8, MXFP8, NVFP4, and block scaling recipes.
transformer_engine/pytorch/module/grouped_linear.py	Added `make_grouped_weights()` method to convert weight parameters into contiguous GroupedTensor storage. Weights are copied and re-registered to share underlying storage.
transformer_engine/common/recipe/init.py	Changed recipe type-checking methods from instance methods to classmethods, using `issubclass()` instead of `isinstance()`. This aligns with `_get_compatible_recipe()` returning class types.
tests/pytorch/test_grouped_tensor.py	New comprehensive test file with 430 lines covering GroupedTensor construction, splitting, quantization for all supported recipes, and verification of contiguous memory layout.

Sequence Diagram

sequenceDiagram
    participant User
    participant GroupedLinear
    participant GroupedTensor
    participant Quantizer
    participant Storage

    Note over User,Storage: Initialization Phase
    User->>GroupedLinear: __init__(num_gemms, in_features, out_features)
    GroupedLinear->>GroupedLinear: register_parameter(weight0...weightN)
    GroupedLinear->>GroupedLinear: reset_parameters()
    GroupedLinear->>GroupedLinear: make_grouped_weights()
    
    Note over GroupedLinear,Storage: Weight Consolidation
    GroupedLinear->>Quantizer: _get_weight_quantizers()
    Quantizer-->>GroupedLinear: [quantizer0...quantizerN]
    GroupedLinear->>GroupedTensor: make_grouped_tensor(num_tensors, shapes, quantizers)
    
    Note over GroupedTensor,Storage: Allocate Contiguous Storage
    GroupedTensor->>GroupedTensor: analyze shape patterns
    GroupedTensor->>GroupedTensor: calculate logical_shape, offsets
    GroupedTensor->>Storage: allocate contiguous buffers (data, scale_inv, etc)
    GroupedTensor->>GroupedTensor: split_into_quantized_tensors()
    GroupedTensor-->>GroupedLinear: grouped_weights with quantized_tensors
    
    Note over GroupedLinear: Copy & Re-register Weights
    loop for each weight i
        GroupedLinear->>GroupedTensor: quantized_tensors[i].copy_(weights[i])
        GroupedLinear->>GroupedLinear: register_parameter(weightI, quantized_tensors[i])
    end
    
    Note over User,Storage: Forward Pass
    User->>GroupedLinear: forward(inp, m_splits)
    GroupedLinear->>GroupedLinear: _get_weight_tensors()
    GroupedLinear->>GroupedLinear: prepare quantizers
    GroupedLinear->>GroupedLinear: _GroupedLinear.apply()
    Note over GroupedLinear: All weights share contiguous storage
    GroupedLinear->>GroupedLinear: general_grouped_gemm(weights, inputs)
    GroupedLinear-->>User: output tensor

greptile-apps

_{9 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-01-15T15:02:33Z

transformer_engine/pytorch/module/grouped_linear.py

+        # TODO(ksivamani): Verify correctness of copy for all recipes.
+        with torch.no_grad():
+            for i in range(self.num_gemms):
+                grouped_weights.quantized_tensors[i].copy_(weights[i])


style: check that the copy operation works correctly for all quantization recipes (FP8, MXFP8, NVFP4, block scaling). the TODO comment on line 771 acknowledges this needs verification.

greptile-apps · 2026-01-15T15:02:34Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+        # TODO(ksivaman): (Do we need multiple quantizers?)
+        # Current implementation assumes all tensors have the different quantizers.
+        # instances but effectively the same quantizer.
+        rowwise_usage = quantizers[0].rowwise_usage if not no_quantization else True
+        columnwise_usage = quantizers[0].columnwise_usage if not no_quantization else False


style: check that all quantizers in the group are compatible. the comment acknowledges uncertainty about whether multiple quantizers are needed, but the implementation assumes they're "effectively the same" - mixed quantization schemes could cause issues.

Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com> Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com> Co-authored-by: Zhongbo Zhu <zhongboz@nvidia.com>

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

zhongbozhu · 2026-01-21T22:42:29Z

tests/pytorch/test_grouped_tensor.py

+        shape = [(512, 512) for _ in range(num_tensors)]
+        quantizers = make_quantizers(quantization, num_tensors, shape)
+
+        grouped_tensor = GroupedTensor.make_grouped_tensor(


This is not graph safe for MOE activation inputs. Maybe we should make a make_grouped_tensor_graph_safe API?

zhongbozhu · 2026-01-21T22:45:20Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+    def __init__(
+        self,
+        num_tensors: int,
+        shape: List[Tuple[int, int]],


We cannot do this, shape is supposed to be a tensor on device and we should have a tiny kernel that computes the offsets
Edit:
It's okay for weights where shapes are static, but for moe activation, this needs to be changed.

zhongbozhu · 2026-01-21T22:57:37Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+
+        no_quantization = quantizers is None or len(quantizers) == 0 or quantizers[0] is None
+
+        # TODO(ksivaman): (Do we need multiple quantizers?)


I think we shouldn't have multiple quantizers, multiple quantizers only make sense for recipes like delayed scaling, where quantizer holds amax values. For blockwise current scaling recipes like mxfp8 and nvfp4, quantizer only holds configurations like whether we are doing RHT, then only one quantizer is enough, and it's better for reducing CPU overhead since pybind don't need to convert multiple objects.

Having separate quantizers also doesn't make sense for cuda graph, because having multiple quantizers implicitly assumes splitting the input, which then implicitly assumes that host knows about the shape of input for each expert, which will then break cuda graph.

Agree. One customer has reviewed this PR and commented that having multiple quantizers would introduce CPU overheads when calling quantizer.update_usage/set_usage/... if the number of tensors is very large, for example, in FSDP.

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>

zhongbozhu · 2026-01-30T21:47:52Z

transformer_engine/pytorch/module/grouped_linear.py

+        if defer_init:
+            return
+
+        weights = [getattr(self, f"weight{i}") for i in range(self.num_gemms)]


How about the case where we do fp8 param gather, each weight in this case is already a Float8Tensor?

zhongbozhu · 2026-01-30T22:34:53Z

transformer_engine/common/include/transformer_engine/transformer_engine.h

+/*! \enum Float8BlockScaleTensorFormat
+ *  \brief Data format for an FP8 block-scaled tensor
+ */
+enum class Float8BlockScaleTensorFormat {


this flag has been deprecated, plus that for grouped tensor, because there is no TP, the format is always GEMM_READY

* changes for pytoch extension; but everything seems to be broken probably unrelated to my changes Signed-off-by: Varun Thumbe <vthumbe@nvidia.com> * fix the issues Signed-off-by: Varun Thumbe <vthumbe@nvidia.com> * comment nvte API since Oleg's PR is not merged Signed-off-by: Varun Thumbe <vthumbe@nvidia.com> * test for all cases: Signed-off-by: Varun Thumbe <vthumbe@nvidia.com> * tensor attributes should be set later Signed-off-by: Varun Thumbe <vthumbe@nvidia.com> --------- Signed-off-by: Varun Thumbe <vthumbe@nvidia.com>

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

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

cyanguwa · 2026-02-09T19:10:11Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+
+        quantized_tensors = self.split_into_quantized_tensors()
+        for i in range(self.num_tensors):
+            self.quantizer.update_quantized(tensors[i], quantized_tensors[i], noop_flag=noop_flag)


Shouldn't this be calling the group_quantize from the C side? Why is it still in a loop of num_tensors iterations? Thanks.

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

cyanguwa · 2026-02-09T22:25:17Z

transformer_engine/pytorch/csrc/type_converters.cpp

+                           GetTransformerEngineDType(tensor_offsets.scalar_type()),
+                           getTensorShape(tensor_offsets));
+  }
+


Could we add the with_gemm_swizzled_scales support here too please?

I think for grouped tensor, since there is no TP, we can always assume with_gemm_swizzled_scales = True because current multi-tensor-swizzle doesn't work with grouped tensor anyway.

Although I agree that having a toggle would be good.

cyanguwa · 2026-02-09T22:25:43Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+
+    def split_into_quantized_tensors(
+        self,
+    ) -> List[Union[QuantizedTensorStorage, torch.Tensor]]:


Could this function be extended to supporting any number n of tensors that satisfies "num_tensors is divisible by n"? This would help with reshaping efforts, for example, in attention.

cyanguwa · 2026-02-09T22:30:53Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+from .nvfp4_tensor_storage import NVFP4TensorStorage
+
+
+class GroupedTensor:


Should this be GroupedTensor or GroupedTensorStorage, like with the other low-precision tensor types?

cyanguwa · 2026-02-09T23:07:45Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+            f"GroupedTensor(num_tensors={self.num_tensors}, "
+            f"shape={self.shape}, "
+            f"logical_shape={self.logical_shape}, "
+            f"dtype={self.get_dtype()})"


Should we add the quantizer information here?

cyanguwa · 2026-02-09T23:08:24Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+    - logical_shape provides the conceptual 2D interpretation
+    - All data is stored on device in contiguous layout
+
+    Note: This structure is used only for combined storage of multiple tensors with the same dtype and scaling mode.


This seems to be duplicate to the first sentence in the docstring.

cyanguwa · 2026-02-09T23:10:18Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+        # Offsets for indexing into contiguous 1D layout (OPTIONAL - not needed if all_same_shape())
+        # tensor_offsets[i] = element offset to start of tensor i (cumulative sum of numel for tensors 0..i-1)
+        # Usage: tensor_i_ptr = data.data_ptr() + tensor_offsets[i] * element_size
+        # If None and all_same_shape(): offset[i] = i * M * N (where M, N are common dimensions)


At first glance, these two tensors are confusing. Maybe we could explain the difference as one being the offset on the element level, one at the byte level?

cyanguwa · 2026-02-09T23:12:09Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+    @staticmethod
+    def make_grouped_tensor_with_shapes(
+        num_tensors: int,
+        shape: List[Tuple[int, int]],


Could this be named "shapes" please?

cyanguwa · 2026-02-10T00:16:05Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+        dtype: Optional[torch.dtype] = None,
+    ) -> GroupedTensor:
+        """
+        Create a GroupedTensor for storing multiple weight tensors of the same shape.


Why "weight" tensors?

cyanguwa · 2026-02-10T00:26:53Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+                        )
+                        result.append(tensor_data)
+                    else:
+                        raise RuntimeError("GroupedTensor has no data to split")


The if/elseif/else code seems to be the same for both "offsets being None" or "not being None" cases, so maybe we can combine the two? Similarly, can the "data" part of the split for both "no_quantization" and "quantization" paths be combined, so there's only one "for i in range(self.num_tensors)" loop?

cyanguwa · 2026-02-10T00:42:40Z

tests/pytorch/test_grouped_tensor.py

+]
+
+
+def make_quantizer(quantization: str, num_tensors: int, shape: List[Tuple[int, int]]) -> Quantizer:


We don't need "num_tensors" as an argument here anymore, I think, because we assume all tensors in the group use the same kind of quantizer.

cyanguwa · 2026-02-10T02:48:14Z

transformer_engine/pytorch/tensor/storage/grouped_tensor.py

+        Quantize the GroupedTensor inplace.
+        """
+
+        quantized_tensors = self.split_into_quantized_tensors()


Should "quantized_tensors" here be "self.quantized_tensors"?

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

zhongbozhu · 2026-02-10T18:23:48Z

tests/pytorch/test_grouped_tensor.py

+
+    @pytest.mark.parametrize(
+        "shape",
+        [[(256, 512), (512, 512), (768, 512)], [(512, 512), (512, 512), (512, 512)]],


please add edge cases to the test:

zero tokens at all

0 tokens in the beginning

0 tokens in the end

0 tokens in the middle

zhongbozhu · 2026-02-10T18:31:27Z

transformer_engine/pytorch/csrc/extensions/cast.cpp

+      py::reinterpret_borrow<py::object>(quantizer), first_dims, logical_first_dim,
+      logical_last_dim);
+
+  NVTE_SCOPED_GIL_RELEASE({


let's build a switch here for recipe like split_quantize

zhongbozhu · 2026-02-10T18:32:33Z

transformer_engine/pytorch/csrc/extensions/pybind.cpp

  m.def("dequantize", &transformer_engine::pytorch::dequantize, "Dequantize", py::arg("input"),
        py::arg("otype"));
-
+  m.def("group_quantize", transformer_engine::pytorch::group_quantize, py::arg("tensor"),


for future grouped quantization for weights, let's add a slot for an optional noop tensor for cuda graph

vthumbe1503 · 2026-02-10T18:35:30Z

transformer_engine/pytorch/csrc/quantizer.cpp

+  }
+
+  if (columnwise_usage) {
+    columnwise_data = at::empty({total_elements}, uint8_opts);


A lot of the code is duplicated throughout for each and e very quantizer. Maybe we should have common code in the base Quantizer implementation

Or maybe call the utility function from type_conerters.cpp.

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Python GroupedTensor and contiguous weights for GroupedLinear

5e39835

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

ksivaman requested review from vthumbe1503 and zhongbozhu January 15, 2026 14:58

ksivaman marked this pull request as draft January 15, 2026 14:58

ksivaman requested a review from ptrendx January 15, 2026 14:58

Merge branch 'main' into grouped_tensor_python

66e7d7f

greptile-apps bot reviewed Jan 15, 2026

View reviewed changes

Graph safe C API for grouped RHT, needs testing

40c619e

Signed-off-by: Zhongbo Zhu <zhongboz@nvidia.com> Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com> Co-authored-by: Zhongbo Zhu <zhongboz@nvidia.com>

ksivaman force-pushed the grouped_tensor_python branch from 2b7ea40 to 40c619e Compare January 16, 2026 07:36

ksivaman added 2 commits January 16, 2026 13:07

Merge branch 'main' into grouped_tensor_python

cf61339

C++ utils, untested

759e7bb

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

zhongbozhu reviewed Jan 21, 2026

View reviewed changes

transformer_engine/pytorch/tensor/storage/grouped_tensor.py Show resolved Hide resolved

Merge branch 'main' into grouped_tensor_python

1d09c2a

Signed-off-by: vthumbe1503 <vthumbe@nvidia.com>

vthumbe1503 mentioned this pull request Jan 23, 2026

MoE training optimization #2438

Open

zhongbozhu reviewed Jan 30, 2026

View reviewed changes

nvMelissa mentioned this pull request Feb 2, 2026

Refactor the grouped functionality to use NVTEGroupedTensor #2510

Open

ksivaman and others added 7 commits February 3, 2026 00:20

Merge branch 'main' into grouped_tensor_python

e1b65ac

Merge branch 'main' into grouped_tensor_python

ebf2194

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

4337520

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

Fix make grouped tensor api

5ab30f5

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Merge branch 'main' into grouped_tensor_python

05dab12

Fixes to tests

68ce836

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

ksivaman mentioned this pull request Feb 6, 2026

[PyTorch] Python GroupedTensor #2654

Open

15 tasks

ksivaman added the MoE label Feb 6, 2026

ksivaman mentioned this pull request Feb 6, 2026

[C] NVFP4 quantization for GroupedTensor #2655

Draft

13 tasks

ksivaman and others added 4 commits February 9, 2026 03:02

Merge branch 'main' into grouped_tensor_python

53c38ec

Fix test

d57651d

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

All tests pass

bd41fd0

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

351b74d

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

cyanguwa reviewed Feb 9, 2026

View reviewed changes

C++ PyTorch GroupedTensor changes WIP

97a1f33

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

cyanguwa reviewed Feb 9, 2026

View reviewed changes

cyanguwa reviewed Feb 10, 2026

View reviewed changes

ksivaman and others added 5 commits February 10, 2026 03:17

Compiles

e1788b3

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

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

9022383

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

Fix runtime failure for test

11095a9

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Fix IMA in mxfp8 GQ

373d9e3

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Add CG test for grouped_quantize

1601960

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

zhongbozhu reviewed Feb 10, 2026

View reviewed changes

vthumbe1503 reviewed Feb 10, 2026

View reviewed changes

Fix recipe tests and FP8 weights

bd57000

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>


		no_quantization = quantizers is None or len(quantizers) == 0 or quantizers[0] is None

		# TODO(ksivaman): (Do we need multiple quantizers?)

		from .nvfp4_tensor_storage import NVFP4TensorStorage


		class GroupedTensor:

		]


		def make_quantizer(quantization: str, num_tensors: int, shape: List[Tuple[int, int]]) -> Quantizer:

[PyTorch] GroupedTensor integration #2600

Are you sure you want to change the base?

[PyTorch] GroupedTensor integration #2600

Conversation

ksivaman commented Jan 15, 2026

Description

Type of change

Changes

Checklist:

Uh oh!

greptile-apps bot commented Jan 15, 2026

Greptile Summary

Key Changes

Implementation Notes

Confidence Score: 4/5

Important Files Changed

Sequence Diagram

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Jan 15, 2026

Choose a reason for hiding this comment

Uh oh!

zhongbozhu Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhongbozhu Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cyanguwa Feb 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cyanguwa Feb 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

[PyTorch] `GroupedTensor` integration #2600

[PyTorch] `GroupedTensor` integration #2600

zhongbozhu Jan 21, 2026 •

edited

Loading

zhongbozhu Jan 21, 2026 •

edited

Loading

cyanguwa Feb 9, 2026 •

edited

Loading

cyanguwa Feb 10, 2026 •

edited

Loading

vthumbe1503 Feb 10, 2026 •

edited

Loading