**Summary**
Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA.
This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend.
**Benchmarking**
To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding.
Settings:
- 1 H100 machine
- `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16`
- dtype `torch.bfloat16`
- `is_causal=False`
- for variable length, we set sequences to be random multiples of 64 up to `max_seq_len`
- 100 runs
| | Variable Length API | SDPA |
|--------|--------------------|----------|
| Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms |
| TFLOPs | 231.812 | 320.840 |
The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length.
**Testing**
Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA.
**Next steps**
Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics.
(This stack builds on top of #162326)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164502
Approved by: https://github.com/v0i0, https://github.com/drisspg
Summary:
* Add `torch._scaled_grouped_mm_v2` with more functionality and
extensibility for future formats
* Add `torch.nn.functional.scaled_grouped_mm` as public entrypoint
* Test both original and v2 functionality
Test Plan:
```
pytest -svv -k grouped test/test_scaled_matmul_cuda.py
```
Reviewers:
Subscribers:
Tasks:
Tags:
Signed-off-by: Simon Layton <simonlayton@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165154
Approved by: https://github.com/drisspg, https://github.com/danielvegamyhre
**Summary**
Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA.
This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend.
**Benchmarking**
To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding.
Settings:
- 1 H100 machine
- `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16`
- dtype `torch.bfloat16`
- `is_causal=False`
- for variable length, we set sequences to be random multiples of 64 up to `max_seq_len`
- 100 runs
| | Variable Length API | SDPA |
|--------|--------------------|----------|
| Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms |
| TFLOPs | 231.812 | 320.840 |
The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length.
**Testing**
Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA.
**Next steps**
Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics.
(This stack builds on top of #162326)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164502
Approved by: https://github.com/v0i0, https://github.com/drisspg
for pipeline parallel, we can have multiple FSDP roots (chunks)
```
model = nn.Sequential([chunk0, chunk1])
fully_shard(model.chunk0)
fully_shard(model.chunk1)
```
we can call `share_comm_ctx` to share all-gather, reduce-scatter, all-reduce cuda streams. this avoids inter-stream memory fragmentation
```
from torch.distributed.fsdp import share_comm_ctx
share_comm_ctx([model.chunk0, model.chunk1])
```
unit test: `pytest -s test/distributed/_composable/fsdp/test_fully_shard_training.py -k test_share_comm_context`
Summary:
Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165024
Approved by: https://github.com/mori360
This PR introduces a way to compile a region of FX graph using `fx.traceback.annotate`.
### UX
1) In the user code, mark the region that you want to be compiled with inductor using `with fx_traceback.annotate({"compile_with_inductor": 0})`. As of now, we just rely on the string `compile_with_inductor` and ignore the integer. As the needs arise, we can update the logic.
Example
```
def fn(x, y):
sin = torch.sin(x)
with fx_traceback.annotate({"compile_with_inductor": 0}):
mul = sin * y
add = mul + 1
return torch.sin(add)
```
2) You have to instruct the compiler to use the annotations with `compile_fx_annotated_nodes_with_inductor` transformation. This is somewhat controversial, and a user might expect that just setting annotation is enough. But for now to control the blast radius, we need to explicitly do this. One such example is
```
# Set the fw and bw compiler of aot_autograd to `compile_fx_annotated_nodes_with_inductor`
def aot_eager_regional_inductor():
return aot_autograd(
fw_compiler=compile_fx_annotated_nodes_with_inductor,
bw_compiler=compile_fx_annotated_nodes_with_inductor,
)
```
3) Fixable in short-term - You have to wrap the user code in `torch.fx.traceback.preserve_node_meta` to ensure that annotations are propagated to the compiler. This is fixable, just need to make CI happy.
### Implementation
1) Relies on `CapabilityBasedPartitioner` to "scoop" out regions based on annotations, and then create subgraphs in the main graph.
2) Call `torch._inductor.standalone_compile` on these subgraphs, and jam the returned callable into the FX graph at the place of call_module
Resulting graph looks something like this - search for `torch__inductor_standalone_compile_inner`
Forward graph
```
class GraphModule(torch.nn.Module):
def forward(self, primals_1: "f32[10]", primals_2: "f32[10]"):
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
sin: "f32[10]" = torch.ops.aten.sin.default(primals_1)
# No stacktrace found for following nodes
inner = torch__inductor_standalone_compile_inner(sin, primals_2)
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:68 in fn, code: add = mul + 1
getitem: "f32[10]" = inner[0]; inner = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:70 in fn, code: return torch.sin(add)
sin_1: "f32[10]" = torch.ops.aten.sin.default(getitem)
return (sin_1, primals_1, primals_2, sin, getitem)
```
Backward graph
```
class GraphModule(torch.nn.Module):
def forward(self, primals_1: "f32[10]", primals_2: "f32[10]", sin: "f32[10]", add: "f32[10]", tangents_1: "f32[10]"):
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
cos_1: "f32[10]" = torch.ops.aten.cos.default(primals_1); primals_1 = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:70 in fn, code: return torch.sin(add)
cos: "f32[10]" = torch.ops.aten.cos.default(add); add = None
mul_1: "f32[10]" = torch.ops.aten.mul.Tensor(tangents_1, cos); tangents_1 = cos = None
# No stacktrace found for following nodes
inner = torch__inductor_standalone_compile_inner(mul_1, sin, primals_2); mul_1 = sin = primals_2 = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:67 in fn, code: mul = sin * y
getitem: "f32[10]" = inner[0]
getitem_1: "f32[10]" = inner[1]; inner = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
mul_4: "f32[10]" = torch.ops.aten.mul.Tensor(getitem_1, cos_1); getitem_1 = cos_1 = None
return (mul_4, getitem)
```
### Some issue raised in the HOP meeting
1) CSE will not differentiate different meta custom nodes and do wrong thing.
2) SAC - The recomputed forward will be smaller than the forward. Will we compile a smaller region than?
3) What happens if you have a op in the middle which does not disturb the topology, is it still 1 subgraph?
4) What happens with the nesting of `fx_traceback.annotate`? Are there any ordering requirements?
5) What are we going to use the annotations for?
a) compile flex
b) streams
c) nn.Module info to organize MoE components for pipelining
d) PP stages
e) Rename graph nodes for more debugging
f) No nested regional compile
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164776
Approved by: https://github.com/SherlockNoMad
ghstack dependencies: #165188
# Motivation
Aligned with other backends, this PR introduces a new API `torch.xpu.is_tf32_supported`, which should be used before `torch.backends.mkldnn.allow_tf32=True` or provide hardware capability information to the Triton
# Additional Context
On Intel Xe architecture and newer, TF32 operations can be accelerated through DPAS (Dot Product Accumulate Systolic) instructions. Therefore, TF32 support can be determined by checking whether the device supports subgroup matrix multiply-accumulate operations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163141
Approved by: https://github.com/EikanWang
This PR introduces a way to compile a region of FX graph using `fx.traceback.annotate`.
### UX
1) In the user code, mark the region that you want to be compiled with inductor using `with fx_traceback.annotate({"compile_with_inductor": 0})`. As of now, we just rely on the string `compile_with_inductor` and ignore the integer. As the needs arise, we can update the logic.
Example
```
def fn(x, y):
sin = torch.sin(x)
with fx_traceback.annotate({"compile_with_inductor": 0}):
mul = sin * y
add = mul + 1
return torch.sin(add)
```
2) You have to instruct the compiler to use the annotations with `compile_fx_annotated_nodes_with_inductor` transformation. This is somewhat controversial, and a user might expect that just setting annotation is enough. But for now to control the blast radius, we need to explicitly do this. One such example is
```
# Set the fw and bw compiler of aot_autograd to `compile_fx_annotated_nodes_with_inductor`
def aot_eager_regional_inductor():
return aot_autograd(
fw_compiler=compile_fx_annotated_nodes_with_inductor,
bw_compiler=compile_fx_annotated_nodes_with_inductor,
)
```
3) Fixable in short-term - You have to wrap the user code in `torch.fx.traceback.preserve_node_meta` to ensure that annotations are propagated to the compiler. This is fixable, just need to make CI happy.
### Implementation
1) Relies on `CapabilityBasedPartitioner` to "scoop" out regions based on annotations, and then create subgraphs in the main graph.
2) Call `torch._inductor.standalone_compile` on these subgraphs, and jam the returned callable into the FX graph at the place of call_module
Resulting graph looks something like this - search for `torch__inductor_standalone_compile_inner`
Forward graph
```
class GraphModule(torch.nn.Module):
def forward(self, primals_1: "f32[10]", primals_2: "f32[10]"):
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
sin: "f32[10]" = torch.ops.aten.sin.default(primals_1)
# No stacktrace found for following nodes
inner = torch__inductor_standalone_compile_inner(sin, primals_2)
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:68 in fn, code: add = mul + 1
getitem: "f32[10]" = inner[0]; inner = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:70 in fn, code: return torch.sin(add)
sin_1: "f32[10]" = torch.ops.aten.sin.default(getitem)
return (sin_1, primals_1, primals_2, sin, getitem)
```
Backward graph
```
class GraphModule(torch.nn.Module):
def forward(self, primals_1: "f32[10]", primals_2: "f32[10]", sin: "f32[10]", add: "f32[10]", tangents_1: "f32[10]"):
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
cos_1: "f32[10]" = torch.ops.aten.cos.default(primals_1); primals_1 = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:70 in fn, code: return torch.sin(add)
cos: "f32[10]" = torch.ops.aten.cos.default(add); add = None
mul_1: "f32[10]" = torch.ops.aten.mul.Tensor(tangents_1, cos); tangents_1 = cos = None
# No stacktrace found for following nodes
inner = torch__inductor_standalone_compile_inner(mul_1, sin, primals_2); mul_1 = sin = primals_2 = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:67 in fn, code: mul = sin * y
getitem: "f32[10]" = inner[0]
getitem_1: "f32[10]" = inner[1]; inner = None
# File: /data/users/anijain/pytorch2/test/dynamo/test_regional_inductor.py:64 in fn, code: sin = torch.sin(x)
mul_4: "f32[10]" = torch.ops.aten.mul.Tensor(getitem_1, cos_1); getitem_1 = cos_1 = None
return (mul_4, getitem)
```
### Some issue raised in the HOP meeting
1) CSE will not differentiate different meta custom nodes and do wrong thing.
2) SAC - The recomputed forward will be smaller than the forward. Will we compile a smaller region than?
3) What happens if you have a op in the middle which does not disturb the topology, is it still 1 subgraph?
4) What happens with the nesting of `fx_traceback.annotate`? Are there any ordering requirements?
5) What are we going to use the annotations for?
a) compile flex
b) streams
c) nn.Module info to organize MoE components for pipelining
d) PP stages
e) Rename graph nodes for more debugging
f) No nested regional compile
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164776
Approved by: https://github.com/SherlockNoMad
Summary:
* Add `torch.nn.functional.scaled_mm` as an abstraction around the C++
methods
* Wraps `torch._scaled_mm_v2` API by default, but user can force use of
the older `torch._scaled_mm` interface.
* Scaled MM tests now run on the new API
Test Plan:
`pytest test/test_scaled_matmul_cuda.py`
Reviewers:
Subscribers:
Tasks:
Tags:
Signed-off-by: Simon Layton <simonlaytonmeta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164142
Approved by: https://github.com/drisspg
ghstack dependencies: #164141
## Summary
- add a CuBLASReductionOption enum so the CUDA context can track reduced-precision and split-K options
- extend the Python bindings, backend helpers, and docs to accept an optional allow_splitk argument for fp16/bf16 matmul controls
- update cuBLAS/cuBLASLt call sites plus dynamo guards and tests to respect the new combinations
## Testing
- python test/test_cuda.py TestCuda.test_cublas_allow_fp16_reduced_precision_reduction_get_set -v *(fails: ModuleNotFoundError: No module named 'psutil')*
------
https://chatgpt.com/codex/tasks/task_e_68e404623178832f8a3e1d34e1e175da
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164766
Approved by: https://github.com/malfet, https://github.com/albanD
Summary:
1\ Certain checkpoint load use cases are not aware of the properties of the data/tensors they want to load.
2\ These usecases include data loader checkpoints, reading data for post processing (when the original model definition is not available).
3\ There, we have to use saved checkpoint (metadata) as our source of truth.
4\ This RFC proposal exposes the checkpoint metadata using a public API.
In this proposal we expose the stored state-dict metadata (minus associated storage/chunk metadata).
Chunk/storage details should not be exposed to the users and is a impl detail of the storage writer/reader.
Test Plan:
UT.
Rollback Plan:
Differential Revision: D80231457
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160610
Approved by: https://github.com/saumishr
Low-level PyTorch APIs should be usable/stable enough at this point but we might move the underlying driver API usage a bit from here...
Built on top of @drisspg 's branch
Pull Request resolved: https://github.com/pytorch/pytorch/pull/159104
Approved by: https://github.com/ngimel
Co-authored-by: drisspg <drisspguessous@gmail.com>
Low-level PyTorch APIs should be usable/stable enough at this point but we might move the underlying driver API usage a bit from here...
Built on top of @drisspg 's branch
Pull Request resolved: https://github.com/pytorch/pytorch/pull/159104
Approved by: https://github.com/ngimel
Co-authored-by: drisspg <drisspguessous@gmail.com>
Summary:
This diff adds the feature of allocating a large pinned memory segment upfront based on the provided config. This large segment is then used to serve all the small pinned memory requests to avoid expensive device level APIs (slow paths).
Example:
PYTORCH_CUDA_ALLOC_CONF=pinned_reserve_segment_size_mb:2048
This reserves a 2GB pinned memory segment for the process and then all incoming small requests are just served from this segment and no cudaHostAlloc/cudaHostRegister apis are being called.
Differential Revision: D83779074
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164501
Approved by: https://github.com/yangw-dev
Previous work #158352 delivered CUDAGraph memory footprint reduction with no replay-time impact, but capture time regressed (up to 20× slower) due to repeated full-graph traversals. See previous benchmark results [here](https://github.com/pytorch/pytorch/pull/158352#issuecomment-3215947565)
This PR removes capture/reply overhead while preserving the memory savings:
1. **Terminals as free markers**
We stop inserting empty nodes and instead record the current stream terminals as free markers. This avoids mutating the user’s graph and keeps semantics unchanged.
2. **Incremental, cached reachability**
We add a **per-graph reuse context** that caches reverse-traversal state:
* `graph_reuse_context[graph].visited[stream]` tracks nodes already seen from that stream’s terminal frontier.
* On each allocation during capture, we resume traversal from the latest terminals and only visit unseen nodes.
* A block is freed when all its recorded markers are in the visited set of its allocation stream—i.e., all markers are proven predecessors of future work.
See [the performance results here](https://docs.google.com/spreadsheets/d/e/2PACX-1vRPvdd9Xa8W87ixbiA0da_qvOhrUAjUpFz0G-_j-MsDnoeRyhEa4_ut_W3rqcg1VVZVFJ-gucwov-3b/pubhtml?gid=1468302443&single=true), we sweep synthetic multi-stream CUDA Graphs built by `capture_benchmark.py` (same as before, we generate random interleaving of alloc/free/join with given probabilities, see [gist here](https://gist.github.com/eee4017/e2092d215b1d4bd46534148939af39e3)), and we compare median capture/replay times and memory. On an NVIDIA H100 PCIe across 24 configs, the optimization preserves reserved memory reduction at ~24–98%, leaves allocated memory unchanged, and brings capture time back to baseline (range 0.96–1.04× vs. baseline) with replay time unchanged (range 0.97–1.11×).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162186
Approved by: https://github.com/eqy, https://github.com/ngimel
