[PGNCCL] Launch kernel on current stream & remove record_stream entirely
#148467
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…147820) Summary: PTD current workflow: - PTD creates its own dedicated `ncclStream` for comm operation - it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us). This diff: - async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead - async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready - pass down async from c10d down to NCCL-PG this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%** Test Plan: - # AMD > before ``` [cenzhao@devgpu039.atn3 ~/fbsource/fbcode (2265d32f0)]$ buck2 run @//mode/opt-amd-gpu -c fbcode.split-dwarf=True //param_bench/train/comms/pt:launcher -- --launcher mpi --nnode 1 --collective all_reduce --b 20M --e 20M --data-type bfloat16 --backend nccl --n 100 --w 5 --envs "NCCL_DEBUG_FILE=/tmp/dedicated_log_rccl.%h.%p.log;NCCL_DEBUG=INFO;NCCL_DEBUG_SUBSYS=INIT,COLL;MSCCL_ALGO_DIR=/data/users/${USER}/fbsource/third-party/rccl/develop/tools/msccl-algorithms;RCCL_MSCCLPP_THRESHOLD=$((128*1024*1024));RCCL_MSCCLPP_ENABLE=1;TORCH_NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK=1;" --size-start-profiler 20M ``` https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree/traces/dynocli/devgpu039.atn3.facebook.com/rank-0.Feb_24_16_19_28.354787.pt.trace.json.gz&bucket=hpc_traces {F1975408857} - c10d::allreduce_(69us) - cudaStreamSync (23us) - nccl::all_reduce(26us) > after ``` [cenzhao@devgpu039.atn3 ~/fbsource/fbcode (2265d32f0)]$ buck2 run @//mode/opt-amd-gpu -c fbcode.split-dwarf=True //param_bench/train/comms/pt:launcher -- --launcher mpi --nnode 1 --collective all_reduce --b 20M --e 20M --data-type bfloat16 --backend nccl --n 100 --w 5 --envs "NCCL_DEBUG_FILE=/tmp/dedicated_log_rccl.%h.%p.log;NCCL_DEBUG=INFO;NCCL_DEBUG_SUBSYS=INIT,COLL;MSCCL_ALGO_DIR=/data/users/${USER}/fbsource/third-party/rccl/develop/tools/msccl-algorithms;RCCL_MSCCLPP_THRESHOLD=$((128*1024*1024));RCCL_MSCCLPP_ENABLE=1;TORCH_NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK=1;TORCH_NCCL_USE_CURRENT_STREAM_AS_NCCL_STREAM=1" --size-start-profiler 20M ``` https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree/traces/dynocli/devgpu039.atn3.facebook.com/rank-4.Feb_24_16_22_56.534269.pt.trace.json.gz&bucket=hpc_traces {F1975408962} - c10d:allreduce_(37us) - cudaStreamSync (gone) - nccl::all_reduce(20us) # NV > before ``` [cenzhao@devgpu019.prn3 ~/fbsource/fbcode (e3f64263c)]$ buck2 run @//mode/opt -c fbcode.split-dwarf=True //param_bench/train/comms/pt:launcher -- --launcher mpi --nnode 1 --collective all_reduce --b 20M --e 20M --data-type bfloat16 --backend nccl --n 100 --w 5 --envs "NCCL_DEBUG_FILE=/tmp/dedicated_log_rccl.%h.%p.log;NCCL_DEBUG=INFO;NCCL_DEBUG_SUBSYS=INIT,COLL;" --size-start-profiler 20M ``` https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree/traces/dynocli/devgpu019.prn3.facebook.com/rank-2.Feb_25_11_11_28.3328768.pt.trace.json.gz&bucket=hpc_traces {F1975437097} - c10d::allreduce_ (62us) - cudaStreamWait (0us) - nccl::all_reduce (47us) > after ``` [cenzhao@devgpu019.prn3 ~/fbsource/fbcode (e3f64263c)]$ buck2 run @//mode/opt -c fbcode.split-dwarf=True //param_bench/train/comms/pt:launcher -- --launcher mpi --nnode 1 --collective all_reduce --b 20M --e 20M --data-type bfloat16 --backend nccl --n 100 --w 5 --envs "NCCL_DEBUG_FILE=/tmp/dedicated_log_rccl.%h.%p.log;NCCL_DEBUG=INFO;NCCL_DEBUG_SUBSYS=INIT,COLL;TORCH_NCCL_USE_CURRENT_STREAM_AS_NCCL_STREAM=1" --size-start-profiler 20M ``` https://www.internalfb.com/intern/perfdoctor/trace_view?filepath=tree/traces/dynocli/devgpu019.prn3.facebook.com/rank-4.Feb_25_11_17_05.3469865.pt.trace.json.gz&bucket=hpc_traces {F1975437192} - c10d::allreduce_ (62us) - cudaStreamWait (gone) - nccl:all_reduce (53us) Differential Revision: D70135605
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/148467
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New FailuresAs of commit 69042db with merge base 10ffd94 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
eqy
reviewed
Mar 4, 2025
| } | ||
|
|
||
| // Therefore, we warn and fall back to the typical recordStream logic. | ||
| // TODO( kwen2501 ): revisit this when we have a better solution. |
There was a problem hiding this comment.
A few options here:
- Change the requirement to be that the user has to call
.wait()even on isend calls. - Add some logic to the destructor that does the
synchronizeStreamor something equivalent to block the destruction on the p2p op
|
Moving over to #148590 as internal user needs |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR has multiple changes to
ProcessGroupNCCL(which unfortunately have to be atomic):cudaStreamWaitEventin PGNCCL #146881work.wait()called by distributed_c10d.py.record_streamand use CPU-side stashing for managing tensor lifetime against recycling.record_streamin c10d causes FSDP2 to over-allocate GPU memory #147168work.wait(), we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion here.Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o