I'm cleaning this PR up as a proper way of disabling functionalization via config in AOTDispatcher. I removed the non-functionalization related changes from the original version:
(1) preventing proxy mode (and functionalization) from incorrectly decomposing CIA ops (Ed has a PR for it here: https://github.com/pytorch/pytorch/pull/164939)
(2) preventing python-dispatcher-based decomps above autograd from running. I'm not doing this for now, will likely do it in a followup
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164577
Approved by: https://github.com/ezyang
ghstack dependencies: #165372
The commits f4d8bc46c7706f872abcb4ec41f0b32207d5d826 added TF32 support for x86 CPUs,
which causes build failures on PowerPC systems with mkldnn.
This patch disables TF32 paths on PowerPC while keeping x86 TF32 support intact,
allowing PyTorch to build successfully on PowerPC.
I have run the mkldnn test case on PowerPC, and it passed successfully.
`pytest test/test_mkldnn.py
87 passed, 2 skipped in 1709.02s (0:28:29`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163454
Approved by: https://github.com/jgong5, https://github.com/malfet
Summary:
Currently `get_c2_fbandroid_xplat_compiler_flags()` is reading the `caffe2.strip_glog` buckconfig which we want to get rid of.
This diff removes the `fbandroid_compiler_flags` arg and merges it with compiler_flags with a nested select and the select version of the method
The goal is to get rid of all the usages of `get_c2_fbandroid_xplat_compiler_flags()` so that we can get rid of the `caffe2.strip_glog` buckconfig
Test Plan: CI
Differential Revision: D84626885
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165558
Approved by: https://github.com/malfet
We have an issue when using fx_traceback.annotate and HOPs that trace joint graphs. HOPs have bodies that have already been traced by Dynamo, and after Animesh's PR, does have the annotations. But when we lower that Dynamo HOP body to aten in either pre-dispatch or post-dispatch, we need to propagate the annotations to the aten nodes.
AOTAutograd does this indirectly by piggybacking off the `PropagateUnbackedSymInts` fx.Interpreter. I'm not sure if all HOPs should be using it to trace their joints or not. This PR adds an interpreter to local_map's implementation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165336
Approved by: https://github.com/yushangdi
**Summary**
The load-balancing problem can be modeled as [identical-machines scheduling](https://en.wikipedia.org/wiki/Identical-machines_scheduling) problem. We already provided an easy-to-extend interface in #161062 for
implementing load-balancing and in this PR we start with adding a Round-Robin solution as an example
and also a verification. This can be easily adapted to other solutions like Shortest-processing-time-first/
Longest-processing-time-first with extra padding added for collectives.
- Added a new type of `_LoadBalancer` implementation `_PTRRLoadBalancer` which is designed for
`flex_attention()`. This load-balance strategy analyzes the `BlockMask` sparsity info and perform
Round-Robin (unlike traditional Round-Robin doing it in circular order, we do in zig-zag order).
- Make `_context_parallel_buffers` and `context_parallel_unshard` handle batched load-balance
index (previously it can only handle non-batched load-balance index), like in `create_cp_block_mask`.
**Test**
`pytest test/distributed/tensor/test_attention.py`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163617
Approved by: https://github.com/fegin
Not sure what exactly we want to have in the message, but that's easy to adjust. I tried to find a reliable test to reproduce this message (happens only when a guard fails right after it's created), but I ended up mocking a `guard_manager.check` function to return `False` to trigger this behavior. I think that's fine, because any other case that we pick (like datetime.now()), we want to patch one day anyway, so every time we make the next patch, will need to chase for another repro test
@williamwen42
Fixes#164990
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165242
Approved by: https://github.com/williamwen42
Summary:
Moves the function used to load CuTeDSL Jinja templates up one level out of the flex attention folder. This way it can be used for more generate Inductor templates in the future.
Test Plan: `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8`
Reviewed By: drisspg
Differential Revision: D84527470
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165576
Approved by: https://github.com/jananisriram
Changed the implementation from an output-based approach to an input-based one to remove `atomicAdd` operations, and it appears to deliver at least a 20× speedup.
The changes are from Yu-Yun <YuYun.Chang@amd.com>.
# Summary: Refactor of the implementation of the `upsample_bilinear2d_backward` opertion on MI300X/MI325X
- The original "scatter-add" approach
- Each thread, representing an output pixel, scattered gradient contributions to four input pixels, using costly atomic operations on MI300X/MI325X GPUs.
- The new "gather-sum" approach
- Each thread is responsible for a single input pixel and gathers all relevant gradient contributions from a small, calculated region of the output tensor (done by the `compute_output_range` device function).
# Breakdown of the code changes
- Inversion of the parallelization strategy of the kernel function `upsample_bilinear2d_backward_out_frame`
- Originally, the main kernel loop was parallelized over the number of elements in the output gradient tensor (`const size_t o_numel = nc * width2 * height2;`).
- Each thread processed one output pixel.
- The new loop is parallelized over the number of elements in the input gradient tensor (`const size_t i_numel = nc * height1 * width1;`).
- Each thread is responsible for calculating the final gradient for a single input pixel.
- The kernel launch changes accordingly in the function `upsample_bilinear2d_backward_out_cuda_template`.
- Added a device function for calculating the range of output pixels that could have possibly used that the input pixel (`input_pos`) during the forward pass interpolation
- This is essentially the mathematical inverse of the forward pass.
- This function tries to prune a thread's search space so that it only needs to inspect a small, local window of the output tensor.
- Gradient calculation approach switching from "scatter-add" to "gather-sum"
- Scatter-add
- For each output pixel, the thread calculated 4 gradient contributions and use `fastAtomicAdd` 4 times to add these values to 4 different (and potentially highly contended) memory locations in the input gradient tensor.
- Gather-sum
- A thread responsible for one input pixel calls `compute_output_range` to determine the small rectangular region of output pixels that influence the input's final gradient value.
- The thread iterates through this region, and for each output pixel in the regionre, it re-calculates the interpolation weights to determine the exact contribution to its specific input pixel.
- All these contributions are accumulated into a private, per-thread register variable (`accscalar_t grad_sum = 0;`).
- W/o any gloabl memory access, this accumulation is extremely fast.
- When the loops are done, the thread performs a single, direct write (non-atomic) of the final summed gradient to its designated location in global memory (`idata[index] = static_cast<scalar_t>(grad_sum);`).
# Why performance gets boosted
- Analysis of the root cause of performance drop
- Ref. (internal only) - https://amd.atlassian.net/wiki/spaces/~glencao2/pages/1140493327/PyTorch__upsample_bilinear2d_backward
- First and foremost, elimination of the contention of atomic operations
- Many parallel threads called `atomicAdd` frequently attempting to update the exact same memory location in the input gradient tensor at the same time.
- The GPU's memory controler has to serialize these operations, effectively nullifying the benefit of parallel capability at those contention points.
- MI300X/MI325X chiplet-based CDNA 3 architeture amplified the issue.
- When contending threads reside on different XCDs, resolving the atomic operation requires high-latency coherence traffic across the Infinity Fabric interconnect.
- The implementation change eliminates hardware-level serialization and cross-chiplet coherence traffic caused by many `atomicAdd`.
- Improved memory access pattern and locality
- Write coalescing
- The regular sum writes `idata[index] = static_cast<scalar_t>(grad_sum);` can be perfectly coalesced by GPUs.
- Read locality
- Even though there are many (potentially repeated) reads from the output tensor (`static_cast<accscalar_t>(odata[output_idx])`), these are highly cache-friendly, meaning the data for one thread is likely to be in the L1 or L2 cache already due to an access from a neighboring thread.
- Trade-off: computation for memory synchronization
- The recalculation of interpolation weights fits well on high-computational-throughput modern GPUs like MI300X/MI325X.
- Removal of atomic operations avoids expensive memory synchronization.
---
Optimizations of `grid_sampler_2d_backward` will be addressed in a separate PR.
Doc for reference: (internal only) https://amd.atlassian.net/wiki/spaces/~glencao2/pages/1162750701/PyTorch__grid_sampler_2d_backward
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164572
Approved by: https://github.com/jeffdaily
Bucketing a number of smallish improvements:
- Account for bucketing in overlap calculation: if an in-flight collective exists with the same bucket key, reduce new collectives estimated time by its latency time
- Update compute domination so we are ordering based on compute idx, as opposed to compute depth, so we never reorder compute. this makes it a bit easier to reason about memory, and pre-fetching, although we can exploring reordering in the future.
- When we wait on a collective, force all collectives on the same process group as it that were enqueued prior to the collective to wait as well.
Better Memory Handling:
- Pre-fetch limiting - when scheduling collectives for overlap, only pre-fetch up to a certain distance, then schedule off-path collectives (which are typically memory reducing).
- When we are above peak memory, schedule waits.
TODO:
- for each compute node, we know its original memory in the graph. we could limit pre-fetching that goes across peak memory
- By scheduling off-path collectives for overlap, we reduce memory, but if there weren't enough compute for overlap, we need to proactively schedule them. not an issue yet on examples.
- config some hard coded constants, clean up enablement (can do in subsequent pr)
On small llama 2d backward :
578 of 618 potentially hideable collectives hidden
original mem 14.4GB, rescheduled mem, 15.9GB
on forward:
254/256 potentially hideable collectives hidden
original mem 5.8 gb, reshceduled mem 5.8GB
WIP: adding tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165318
Approved by: https://github.com/ezyang, https://github.com/IvanKobzarev
ghstack dependencies: #164738, #164783, #164944, #164945, #165059
For `test_graph_partition_with_memory_plan_reuse`, before this PR, when using graph partition, it would error ([P1992728479](https://www.internalfb.com/phabricator/paste/view/P1992728479)):
```
def partition_0(args):
...
del buf0
return (buf3, buf4, buf5, buf2, primals_4, )
...
File "/tmp/torchinductor_boyuan/ww/cwwc7ukfqscg2vy6ankby2fizdb377tvgyx3fwdgddrxe3g47jg6.py", line 132, in partition_0
return (buf3, buf4, buf5, buf2, primals_4, )
^^^^
NameError: name 'buf2' is not defined. Did you mean: 'buf0'?
```
When not using graph partition, it would work and give the following code ([P1992997521](https://www.internalfb.com/phabricator/paste/view/P1992997521)):
```
def call(self, args):
...
buf2 = buf0; del buf0 # reuse
...
```
Note that the issue is buf0 is not reused for buf2 when using graph partition.
Why? Because the codegen runs `run_wrapper_ir_passes` and `memory_plan_reuse`, which pops tailing `MemoryPlanningLine` unless it is in graph output by checking `V.graph.get_output_names()`. However, for graph partition, we should check the output of the current partition instead of the graph before partition.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165514
Approved by: https://github.com/ProExpertProg, https://github.com/eellison
