ezyang
Follow
Pinned
4,395 contributions in the last year
Activity overview
Contribution activity
June 2023
Created 58 commits in 5 repositories
Created 2 repositories
- ezyang/transformers Python
- ezyang/pdfs
Created a pull request in pytorch/pytorch that received 50 comments
CUDA graphs overrides dynamic shapes and forces specialization
Stack from ghstack (oldest at bottom): -> #103290 Previously, cudagraphs and dynamic_shapes were incompatible and enabling dynamic shapes would f…
+142
−33
•
50
comments
Opened 76 other pull requests in 5 repositories
pytorch/pytorch
14
open
58
closed
- [TESTING] Enable automatic_dynamic_shapes by default
- cm3leon_generate is at edge of timeout, so bump it up
- Add automatic_dynamic_shapes test configuration
- Switch dynamic_shapes to True by default
- Delete ifdyn and ifunspec combinators
- Continue simplifying dynamic shapes tests
- Refactor tests for dynamic shapes
- When patching test class, don't run the original tests
- [TEST] Enable automatic dynamic
- Always register SHAPE_ENV guard
- [FAILING] Amend requires_static_shapes conditions
- [TEST] Dynamic default
- Don't test dynamic_shapes in tensor_always_has_static_shape
- Don't test dynamic_shapes in profiler
- Detect symbolic tracing_mode with free_symbols
- [FAILING] Make good on promised TODO code cleanup
- Enable Python dispatcher when ShapeProp with fake mode
- Use free_symbols to determine if convolutions involve dynamic shapes
- indexing_dtype_strength_reduction more aggressive free_symbols tests
- Make all CI commit pin changes trigger ciflow/inductor
- Rewrite size/stride/numel TensorVariable handling
- Make specialized attributes on Tensor mandatory
- Use assertEqual() instead of assertTrue(same())
- Remove numpy_ndarray_as_tensor special case
- Add big doc to wrap_fx_proxy_cls
- Some pull requests not shown.
pytorch/test-infra
1
merged
pytorch/benchmark
1
open
openai/triton
1
merged
huggingface/transformers
1
merged
Reviewed 60 pull requests in 4 repositories
pytorch/pytorch
25 pull requests
- [dynamo][numpy] Add support for builtin functions
- Refactor tests for dynamic shapes
- Improve DDPOptimizer Logging
- Use free_symbols to determine if convolutions involve dynamic shapes
- Move locals/globals to output graph, make it easier to access them anywhere
- Use CUDA DSA in caffe2/operators
- Stop disabling ShapeProp with dynamic_shapes for mkldnn
- Simple Source traversal util
- [dynamo][numpy] Install numpy_pytorch_interop in ci jobs
- Strengthen partially supported invariant of base for chained sources
- Rewrite size/stride/numel TensorVariable handling
- Add big doc to wrap_fx_proxy_cls
- Make specialized attributes on Tensor mandatory
- [c10d] Add xpu to the default device supported by user specified backend
- [PyTorch] Redirect c10::optional to std::optional
- [benchmark][compile] Limit number of bounding boxes to 5
-
[pt2] add metas for
avg_pool3dandavg_pool3d_backward -
[pt2] add
SymIntsupport forbilinear -
[pt2] add
SymIntsupport forcosine_similarity - [dynamo][numpy] Support ndarray methods
- Refactor wrap_fx_proxy_cls, no symbolic_shapes test anymore
- Always create ShapeEnv, always apply unspec logic
- Update torchbench pin - torchrec_dlrm moved to canary
- CUDA graphs overrides dynamic shapes and forces specialization
- Compiled Graph Cache for Inductor
- Some pull request reviews not shown.
pytorch/xla
1 pull request
pytorch/benchmark
1 pull request
pytorch/test-infra
1 pull request
Created an issue in pytorch/test-infra that received 4 comments
Not all PR jobs are getting run on main
On https://hud.pytorch.org/pr/99469 I have this config run on my job: linux-bionic-cuda11.8-py3.10-gcc7-sm86 / test (default, 2, 5, linux.g5.4xlarg…
4
comments
Opened 31 other issues in 4 repositories
pytorch/pytorch
22
open
2
closed
- torch._dynamo.exc.Unsupported: call_function BuiltinVariable(all) [ListIteratorVariable()] {} in DynamicShapesReproTests.test_chunk_reformer_ff_dynamic_shapes
- torch._dynamo.exc.InternalTorchDynamoError: SymNodeVariable() is not a constant on DynamicShapesMiscTests.test_slice_input
- torch._dynamo.exc.Unsupported: torch.* op returned non-Tensor bool call_method is_contiguous on DynamicShapesFunctionTests.test_is_contiguous_memory_format_dynamic_shapes
- test_fstrings2 fails with dynamic
- test_builtin_getitem regression
- torch.fx.passes.split_module.split_module doesn't support dynamic shapes
- test_generate_tensor_from_list_of_numpy_primitive_type fails if run under pytest
- test_workspace_allocation_error fails on my local devgpu
- python test/inductor/test_split_cat_fx_passes.py -k test_consecutive_split_merge fails, but running all tests together succeeds
- TORCH_LOGS=guards doesn't print tensor match guards
- breakpoint() in torch.compile region behaves oddly
- Allow overriding __repr__ to call dataclass_repr (infinite recursion right now)
- PT2 AOTAutograd: view of a view which was created in no_grad mode and is being modified inplace with grad mode enabled
- lit-llama lora fine tuning NetworkXUnbounded: Infinite capacity path, flow unbounded above
- RuntimeError: quantile() input tensor must be either float or double dtype
- Preserve weight_g/weight_v accessors on new weight_norm
- Pickling weight_norm does not actually preserve the weight norm
- weight_norm cannot be deepcopy
- After dynamo minifier generates repros that don't entirely match what we minified over
- Inductor: delete code that extracts out sizevars by inspecting tensor inputs to find a size that handled it
- pdb but for dynamo (and time travel debugging)
- Boolean mask setitem with differentiable tensors fails in dynamo
- Dynamo should only unroll loops by a preset factor (unless otherwise explicitly instructed)
- mark_dynamic may error too aggressively