microsoft/mscclpp
1
0
Fork 0
mirror of https://github.com/microsoft/mscclpp.git synced 2026-05-13 09:46:00 +00:00
Code Issues Packages Projects Releases Wiki Activity
935 Commits 104 Branches 16 Tags
2c52937b26e6b72846cb8bec2f7479fb90162913
Commit Graph

7 Commits

Author SHA1 Message Date
Binyang Li
2c52937b26 Fix FP8 ROCm build/test issues and dtype naming (#792) 
## Summary
- Fix ROCm FP8 build failure by using the actual FP8 `DataType` enum
constants in allreduce packet tuning.
- Fix FP8 E4M3FNUZ test encoding so small negative values do not produce
the FNUZ NaN byte (`0x80`).
- Align FP8 `DataType` enum constants and Python bindings with
torch-style names (`FLOAT8_E4M3FN`, `FLOAT8_E4M3FNUZ`, `FLOAT8_E5M2FNUZ`
/ `float8_e4m3fn`, `float8_e4m3fnuz`, `float8_e5m2fnuz`).

## Validation
- `./tools/lint.sh`
- `make -j` from `build/`
- `mpirun --allow-run-as-root -np 8 python3 -m pytest
python/test/test_fp8_accum.py -q` (`36 passed, 9 skipped`)
- `DTYPE=float8_e4m3fnuz ACCUM_DTYPE=float32 torchrun --nnodes=1
--nproc_per_node=8
examples/torch-integration/customized_comm_with_tuning.py`

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-04-28 15:02:22 -07:00
Binyang Li
96a72bbd3e Support E4M3B15 datatype (#765) 
## Summary

- **Add `fp8_e4m3b15` datatype**: A software-defined FP8 type with 4
exponent bits, 3 mantissa bits, and bias=15 (max finite value: 0.9375).
Implemented entirely in software with no HW dependency, using
Triton-style bit manipulation through fp16 as intermediate for efficient
conversion.
- **Add mixed-precision accumulation for allreduce**: All allreduce
algorithm variants (packet, NVLS packet, fullmesh, RSAG zero-copy, and
others) now support a configurable `accumDtype` parameter, enabling FP8
inputs to be reduced in float16 or float32 for higher accuracy.
- **Propagate `accumDtype` through the full API**: The new parameter is
threaded from `Algorithm::execute()` → `NativeAlgorithm` → `KernelFunc`
→ dispatch → CUDA kernels, with `DataType::AUTO` as the default
(resolves to input dtype at runtime).
- **Add FP8 accumulation correctness tests**: New `test_fp8_accum.py`
validates that higher-precision accumulation produces results at least
as accurate as native FP8 accumulation across multiple algorithms and
sizes. Skipped on CUDA SM < 89 (pre-Hopper); runs on HIP/ROCm.
- **Add `test_fp8_accum.py` to CI**: Azure Pipeline `ut.yml` now runs
FP8 accumulation tests alongside existing pytests.
- **NCCL shim logging cleanup**: Migrated `printf`-style `WARN`/`INFO`
calls to streaming-style logging.

## Key files

| Area | Files |
|------|-------|
| New datatype + vector ops | `include/mscclpp/gpu_data_types.hpp` |
| Accumulation reduce helpers | `src/core/include/reduce_kernel.hpp` |
| Algorithm API (`accumDtype`) | `include/mscclpp/algorithm.hpp`,
`src/core/algorithm.cc` |
| Allreduce kernels | `src/ext/collectives/allreduce/*.cu` |
| Dispatch + common | `src/ext/collectives/include/allreduce/common.hpp`
|
| Python bindings | `python/csrc/algorithm.cpp`,
`python/mscclpp/_core/algorithm.py` |
| Tests | `python/test/test_fp8_accum.py` |
| CI | `.azure-pipelines/templates/ut.yml` |

## Test plan

- [x] CI passes on H100 (CUDA SM 90) — full FP8 E4M3 + E4M3B15
accumulation tests
- [x] CI passes on A100 (CUDA SM 80) — FP8 tests correctly skipped
- [x] CI passes on MI300X (ROCm) — FP8 tests run via HIP
- [x] Existing `test_mscclpp.py` tests continue to pass
- [x] NCCL shim builds and runs correctly with new `accumDtype` defaults

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
 2026-04-07 13:37:02 -07:00
Binyang Li
bf946ea51e Fix multicast handle leak, cuMemMap offset handling, and rename NVLS allreduce algorithms (#759) 
## Summary

This PR addresses a multicast resource leak, fixes `cuMemMap` offset
handling for multicast handles, renames NVLS allreduce algorithm classes
for clarity, and adds a new unit test for `SwitchChannel`.

### Bug Fixes

#### 1. Fix multicast allocation handle leak in `createMulticast()`
(`gpu_ipc_mem.cc`)

`GpuIpcMemHandle::createMulticast()` called
`cuMulticastCreate(&allocHandle, ...)` but never released the local
`allocHandle` after exporting it to shareable handles (POSIX FD /
Fabric). This caused a reference count leak — the multicast object was
never freed even after all mappings and imported handles were released.

Per the [CUDA Driver API docs for
`cuMemRelease`](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__VA.html):
> *"The memory allocation will be freed when all outstanding mappings to
the memory are unmapped and when all outstanding references to the
handle (including its shareable counterparts) are also released."*

The fix adds `cuMemRelease(allocHandle)` after export, matching the
existing pattern used for regular allocations in
`GpuIpcMemHandle::create()`.

**Impact:** Without this fix, repeated creation/destruction of NVLS
connections causes OOM after ~120 iterations when allocating 1GB
multicast buffers on H100.

#### 2. Fix `cuMemMap` offset for multicast handles (`gpu_ipc_mem.cc`)

`cuMemMap` requires `offset=0` for multicast handles. Previously, the
code attempted to map at a non-zero offset within the multicast object,
leading to errors when binding multiple buffers to the same
`NvlsConnection`. The fix maps the entire range `[0, mcOffset +
bufferSize)` and returns the pointer offset by `mcOffset`. This only
consumes extra virtual address space; no additional physical memory is
used.

### Refactoring

#### 3. Rename NVLS allreduce algorithm classes

Renamed for clarity:
- `AllreduceNvls` → `AllreduceNvlsZeroCopy`
- `AllreduceNvlsWithCopy` → `AllreduceNvlsWarpPipeline`
- `AllreduceNvlsWithCopy2` → `AllreduceNvlsBlockPipeline`

Updated all references in builder, selector, docs, and examples.

#### 4. Move `nvlsConnections` setup to `initialize()`

Moved `nvlsConnections_` from `AlgorithmCtx` (which no longer has this
member) to individual algorithm class members, initialized in their
`initialize()` methods.

### Tests

#### 5. Add `TwoChannelsSameConnection` test

New unit test that creates two `SwitchChannel` instances from the same
`NvlsConnection`, performs reduce operations on both, and verifies
correctness. This exercises the multi-bind path that triggered the
`cuMemMap` offset fix.

### Files Changed

- `src/core/gpu_ipc_mem.cc` — multicast handle leak fix + cuMemMap
offset fix
- `src/ext/collectives/allreduce/allreduce_nvls_zero_copy.cu` (renamed)
- `src/ext/collectives/allreduce/allreduce_nvls_warp_pipeline.cu`
(renamed)
- `src/ext/collectives/allreduce/allreduce_nvls_block_pipeline.cu`
(renamed)
- `src/ext/collectives/allreduce/allreduce_nvls_packet.cu` —
nvlsConnections fix
- `src/ext/collectives/include/allreduce/*.hpp` — renamed headers
- `src/ext/collectives/algorithm_collection_builder.cc` — updated
references
- `src/ext/nccl/algorithm_selector.cc` — updated algorithm names
- `test/mp_unit/switch_channel_tests.cu` — new test
- `docs/guide/mscclpp-torch-integration.md` — updated names
- `examples/torch-integration/customized_comm_with_default_algo.py` —
updated names
 2026-03-09 10:22:45 -07:00
Binyang Li
25435acf5d Add new algos for GB200 (#747) 
- Add new algos (allreduce_rsag, allreduce_rsag_pipeline and
allreduce_rsag_zero_copy) for GB200.
- Add IB stub for non-IB env
- Provides example for algorithm tunning with different nblocks/nthreads

Perf for allreduce_rsag
```
#                                                              out-of-place                       in-place          
#       size         count      type   redop    root     time   algbw   busbw  #wrong     time   algbw   busbw  #wrong 
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)             (us)  (GB/s)  (GB/s)         
     1048576        262144     float     sum      -1    25.16   41.67   62.51       0    23.73   44.18   66.27       0
     2097152        524288     float     sum      -1    26.06   80.47  120.71       0    25.31   82.86  124.29       0
     4194304       1048576     float     sum      -1    31.09  134.93  202.39       0    30.75  136.39  204.58       0
     8388608       2097152     float     sum      -1    45.52  184.29  276.43       0    45.13  185.87  278.80       0
    16777216       4194304     float     sum      -1    75.73  221.53  332.30       0    75.51  222.18  333.27       0
    33554432       8388608     float     sum      -1   137.25  244.48  366.72       0   137.22  244.54  366.81       0
    67108864      16777216     float     sum      -1   271.34  247.32  370.99       0   270.86  247.76  371.65       0
   134217728      33554432     float     sum      -1   534.25  251.22  376.84       0   534.43  251.14  376.71       0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 264.454 
#
# Collective test concluded: all_reduce_perf
```

perf for allreduce_rsag_pipeline
```
#                                                              out-of-place                       in-place          
#       size         count      type   redop    root     time   algbw   busbw  #wrong     time   algbw   busbw  #wrong 
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)             (us)  (GB/s)  (GB/s)         
     1048576        262144     float     sum      -1    61.57   17.03   25.55       0    61.51   17.05   25.57       0
     2097152        524288     float     sum      -1    61.31   34.20   51.31       0    61.23   34.25   51.38       0
     4194304       1048576     float     sum      -1    61.62   68.06  102.10       0    61.84   67.83  101.74       0
     8388608       2097152     float     sum      -1    61.97  135.37  203.06       0    61.89  135.53  203.30       0
    16777216       4194304     float     sum      -1    63.15  265.65  398.48       0    62.89  266.76  400.15       0
    33554432       8388608     float     sum      -1   100.63  333.46  500.19       0    99.76  336.34  504.51       0
    67108864      16777216     float     sum      -1   180.04  372.75  559.13       0   179.75  373.34  560.01       0
   134217728      33554432     float     sum      -1   339.60  395.23  592.84       0   338.16  396.91  595.36       0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 304.665 
#
# Collective test concluded: all_reduce_perf
```

perf for allreduce_rsag_zero_copy
```
#                                                              out-of-place                       in-place          
#       size         count      type   redop    root     time   algbw   busbw  #wrong     time   algbw   busbw  #wrong 
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)             (us)  (GB/s)  (GB/s)         
     1048576        262144     float     sum      -1    14.99   69.93  104.90       0    14.44   72.61  108.92       0
     2097152        524288     float     sum      -1    16.19  129.56  194.33       0    15.85  132.32  198.48       0
     4194304       1048576     float     sum      -1    21.19  197.98  296.97       0    20.64  203.20  304.81       0
     8388608       2097152     float     sum      -1    31.04  270.27  405.41       0    30.68  273.44  410.16       0
    16777216       4194304     float     sum      -1    50.34  333.26  499.89       0    50.15  334.51  501.77       0
    33554432       8388608     float     sum      -1    89.58  374.56  561.84       0    88.65  378.48  567.73       0
    67108864      16777216     float     sum      -1   165.69  405.03  607.54       0   163.64  410.10  615.16       0
   134217728      33554432     float     sum      -1   323.19  415.28  622.93       0   318.01  422.05  633.07       0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 414.619 
#
# Collective test concluded: all_reduce_perf
```

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com>
Co-authored-by: chhwang <8018170+chhwang@users.noreply.github.com>
Co-authored-by: Qinghua Zhou <qinghuazhou@microsoft.com>
Co-authored-by: Caio Rocha <caiorocha@microsoft.com>
 2026-02-24 16:43:23 -08:00
Binyang Li
4701ae3a95 Update dtype name (#748) 
- Change FP8_E4M3/FP8_E5M2 to FLOAT8_E4M3/FLOAT8_E5M2
- Add torch.uint8 to DataType.uint8 mapping
 2026-02-18 10:35:44 -08:00
Qinghua Zhou
edc9c38751 Support uint8 data type for Allreduce (#736) 
Support uint8 data type for Allreduce.
Current limitation: uint8 is not supported for NVLS.

Performance results with RCCL-test with MSCCLPP on MI300X:


\# out-of-place in-place
\# size count type redop root time algbw busbw #wrong time algbw busbw
#wrong
\# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
1024 | 512 | half | sum | -1 | 5.39 | 0.19 | 0.33 | 0 | 5.45 | 0.19 |
0.33 | 0
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
2048 | 1024 | half | sum | -1 | 5.53 | 0.37 | 0.65 | 0 | 5.63 | 0.36 |
0.64 | 0
4096 | 2048 | half | sum | -1 | 5.55 | 0.74 | 1.29 | 0 | 5.56 | 0.74 |
1.29 | 0
8192 | 4096 | half | sum | -1 | 5.8 | 1.41 | 2.47 | 0 | 5.84 | 1.4 |
2.46 | 0
16384 | 8192 | half | sum | -1 | 6.57 | 2.49 | 4.36 | 0 | 6.56 | 2.5 |
4.37 | 0
32768 | 16384 | half | sum | -1 | 8.02 | 4.09 | 7.15 | 0 | 8.06 | 4.07 |
7.11 | 0
65536 | 32768 | half | sum | -1 | 8.77 | 7.47 | 13.07 | 0 | 8.82 | 7.43
| 13 | 0
131072 | 65536 | half | sum | -1 | 9.61 | 13.64 | 23.87 | 0 | 9.78 |
13.4 | 23.45 | 0
262144 | 131072 | half | sum | -1 | 11.68 | 22.44 | 39.27 | 0 | 12.1 |
21.67 | 37.93 | 0
524288 | 262144 | half | sum | -1 | 13.77 | 38.08 | 66.64 | 0 | 13.87 |
37.79 | 66.13 | 0
1048576 | 524288 | half | sum | -1 | 19.11 | 54.87 | 96.03 | 0 | 19.27 |
54.42 | 95.24 | 0
2097152 | 1048576 | half | sum | -1 | 24.1 | 87 | 152.26 | 0 | 24.24 |
86.52 | 151.41 | 0
4194304 | 2097152 | half | sum | -1 | 37.16 | 112.87 | 197.52 | 0 |
37.44 | 112.03 | 196.06 | 0
8388608 | 4194304 | half | sum | -1 | 61.53 | 136.33 | 238.58 | 0 |
61.68 | 135.99 | 237.99 | 0
16777216 | 8388608 | half | sum | -1 | 108.8 | 154.22 | 269.88 | 0 |
109.2 | 153.6 | 268.79 | 0
33554432 | 16777216 | half | sum | -1 | 197.8 | 169.68 | 296.94 | 0 |
198.6 | 168.92 | 295.61 | 0
67108864 | 33554432 | half | sum | -1 | 384.6 | 174.51 | 305.39 | 0 |
385.1 | 174.27 | 304.98 | 0
134217728 | 67108864 | half | sum | -1 | 754.1 | 177.99 | 311.48 | 0 |
754.9 | 177.78 | 311.12 | 0
268435456 | 134217728 | half | sum | -1 | 1491.8 | 179.94 | 314.89 | 0 |
1493.2 | 179.77 | 314.6 | 0
536870912 | 268435456 | half | sum | -1 | 2979.6 | 180.18 | 315.31 | 0 |
2983.9 | 179.92 | 314.87 | 0


\# out-of-place in-place
\# size count type redop root time algbw busbw #wrong time algbw busbw
#wrong
\# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
1024 | 1024 | fp8_e4m3 | sum | -1 | 5.4 | 0.19 | 0.33 | 0 | 5.45 | 0.19
| 0.33 | 0
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
2048 | 2048 | fp8_e4m3 | sum | -1 | 5.5 | 0.37 | 0.65 | 0 | 5.6 | 0.37 |
0.64 | 0
4096 | 4096 | fp8_e4m3 | sum | -1 | 5.61 | 0.73 | 1.28 | 0 | 5.68 | 0.72
| 1.26 | 0
8192 | 8192 | fp8_e4m3 | sum | -1 | 5.96 | 1.38 | 2.41 | 0 | 5.98 | 1.37
| 2.4 | 0
16384 | 16384 | fp8_e4m3 | sum | -1 | 6.49 | 2.52 | 4.42 | 0 | 6.58 |
2.49 | 4.36 | 0
32768 | 32768 | fp8_e4m3 | sum | -1 | 8.09 | 4.05 | 7.09 | 0 | 8.15 |
4.02 | 7.03 | 0
65536 | 65536 | fp8_e4m3 | sum | -1 | 8.58 | 7.64 | 13.37 | 0 | 8.7 |
7.53 | 13.18 | 0
131072 | 131072 | fp8_e4m3 | sum | -1 | 9.44 | 13.88 | 24.29 | 0 | 9.62
| 13.63 | 23.85 | 0
262144 | 262144 | fp8_e4m3 | sum | -1 | 10.12 | 25.9 | 45.32 | 0 | 10.37
| 25.27 | 44.22 | 0
524288 | 524288 | fp8_e4m3 | sum | -1 | 13.73 | 38.19 | 66.82 | 0 |
13.89 | 37.74 | 66.04 | 0
1048576 | 1048576 | fp8_e4m3 | sum | -1 | 18.66 | 56.2 | 98.34 | 0 |
18.92 | 55.41 | 96.97 | 0
2097152 | 2097152 | fp8_e4m3 | sum | -1 | 24.54 | 85.46 | 149.56 | 0 |
24.63 | 85.16 | 149.03 | 0
4194304 | 4194304 | fp8_e4m3 | sum | -1 | 37.79 | 110.98 | 194.21 | 0 |
38.05 | 110.22 | 192.88 | 0
8388608 | 8388608 | fp8_e4m3 | sum | -1 | 62.22 | 134.82 | 235.94 | 0 |
62.63 | 133.94 | 234.4 | 0
16777216 | 16777216 | fp8_e4m3 | sum | -1 | 109.9 | 152.62 | 267.09 | 0
| 110.4 | 151.9 | 265.83 | 0
33554432 | 33554432 | fp8_e4m3 | sum | -1 | 201.1 | 166.82 | 291.94 | 0
| 202.3 | 165.84 | 290.22 | 0
67108864 | 67108864 | fp8_e4m3 | sum | -1 | 390 | 172.06 | 301.11 | 0 |
390.2 | 171.99 | 300.99 | 0
134217728 | 134217728 | fp8_e4m3 | sum | -1 | 763.9 | 175.7 | 307.47 | 0
| 764.2 | 175.62 | 307.34 | 0
268435456 | 268435456 | fp8_e4m3 | sum | -1 | 1509.5 | 177.83 | 311.2 |
0 | 1510.1 | 177.76 | 311.08 | 0
536870912 | 536870912 | fp8_e4m3 | sum | -1 | 3010.2 | 178.35 | 312.11 |
0 | 3014.2 | 178.11 | 311.7 | 0



\# out-of-place in-place
\# size count type redop root time algbw busbw #wrong time algbw busbw
#wrong
\# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
1024 | 1024 | fp8_e5m2 | sum | -1 | 5.41 | 0.19 | 0.33 | 0 | 5.44 | 0.19
| 0.33 | 0
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
2048 | 2048 | fp8_e5m2 | sum | -1 | 5.5 | 0.37 | 0.65 | 0 | 5.67 | 0.36
| 0.63 | 0
4096 | 4096 | fp8_e5m2 | sum | -1 | 5.61 | 0.73 | 1.28 | 0 | 5.69 | 0.72
| 1.26 | 0
8192 | 8192 | fp8_e5m2 | sum | -1 | 5.96 | 1.37 | 2.4 | 0 | 6 | 1.36 |
2.39 | 0
16384 | 16384 | fp8_e5m2 | sum | -1 | 6.63 | 2.47 | 4.32 | 0 | 6.59 |
2.49 | 4.35 | 0
32768 | 32768 | fp8_e5m2 | sum | -1 | 8.07 | 4.06 | 7.1 | 0 | 8.16 |
4.02 | 7.03 | 0
65536 | 65536 | fp8_e5m2 | sum | -1 | 8.62 | 7.61 | 13.31 | 0 | 8.73 |
7.51 | 13.14 | 0
131072 | 131072 | fp8_e5m2 | sum | -1 | 9.43 | 13.9 | 24.33 | 0 | 9.6 |
13.66 | 23.9 | 0
262144 | 262144 | fp8_e5m2 | sum | -1 | 10.11 | 25.94 | 45.39 | 0 |
10.38 | 25.26 | 44.21 | 0
524288 | 524288 | fp8_e5m2 | sum | -1 | 13.73 | 38.19 | 66.84 | 0 |
13.87 | 37.79 | 66.13 | 0
1048576 | 1048576 | fp8_e5m2 | sum | -1 | 18.65 | 56.22 | 98.39 | 0 |
18.93 | 55.38 | 96.92 | 0
2097152 | 2097152 | fp8_e5m2 | sum | -1 | 24.54 | 85.47 | 149.57 | 0 |
24.63 | 85.16 | 149.03 | 0
4194304 | 4194304 | fp8_e5m2 | sum | -1 | 37.84 | 110.83 | 193.96 | 0 |
38.01 | 110.36 | 193.12 | 0
8388608 | 8388608 | fp8_e5m2 | sum | -1 | 62.32 | 134.61 | 235.58 | 0 |
62.55 | 134.12 | 234.71 | 0
16777216 | 16777216 | fp8_e5m2 | sum | -1 | 110 | 152.58 | 267.01 | 0 |
110.3 | 152.12 | 266.21 | 0
33554432 | 33554432 | fp8_e5m2 | sum | -1 | 201.1 | 166.9 | 292.07 | 0 |
201.8 | 166.26 | 290.96 | 0
67108864 | 67108864 | fp8_e5m2 | sum | -1 | 390 | 172.07 | 301.12 | 0 |
390.5 | 171.87 | 300.78 | 0
134217728 | 134217728 | fp8_e5m2 | sum | -1 | 763.9 | 175.69 | 307.46 |
0 | 764.5 | 175.56 | 307.23 | 0
268435456 | 268435456 | fp8_e5m2 | sum | -1 | 1509.4 | 177.84 | 311.22 |
0 | 1509.8 | 177.8 | 311.14 | 0
536870912 | 536870912 | fp8_e5m2 | sum | -1 | 3013 | 178.18 | 311.82 | 0
| 3018 | 177.89 | 311.31 | 0


\# out-of-place in-place
\# size count type redop root time algbw busbw #wrong time algbw busbw
#wrong
\# (B) (elements) (us) (GB/s) (GB/s) (us) (GB/s) (GB/s)
1024 | 1024 | uint8 | sum | -1 | 5.46 | 0.19 | 0.33 | 0 | 5.46 | 0.19 |
0.33 | 0
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
2048 | 2048 | uint8 | sum | -1 | 5.54 | 0.37 | 0.65 | 0 | 5.63 | 0.36 |
0.64 | 0
4096 | 4096 | uint8 | sum | -1 | 5.61 | 0.73 | 1.28 | 0 | 5.63 | 0.73 |
1.27 | 0
8192 | 8192 | uint8 | sum | -1 | 5.9 | 1.39 | 2.43 | 0 | 5.9 | 1.39 |
2.43 | 0
16384 | 16384 | uint8 | sum | -1 | 6.6 | 2.48 | 4.35 | 0 | 6.64 | 2.47 |
4.32 | 0
32768 | 32768 | uint8 | sum | -1 | 8.99 | 3.65 | 6.38 | 0 | 8.99 | 3.64
| 6.38 | 0
65536 | 65536 | uint8 | sum | -1 | 9.44 | 6.94 | 12.15 | 0 | 9.58 | 6.84
| 11.98 | 0
131072 | 131072 | uint8 | sum | -1 | 11.72 | 11.18 | 19.57 | 0 | 11.83 |
11.08 | 19.4 | 0
262144 | 262144 | uint8 | sum | -1 | 12.29 | 21.32 | 37.31 | 0 | 12.45 |
21.05 | 36.84 | 0
524288 | 524288 | uint8 | sum | -1 | 13.87 | 37.8 | 66.15 | 0 | 13.93 |
37.64 | 65.88 | 0
1048576 | 1048576 | uint8 | sum | -1 | 19.11 | 54.88 | 96.04 | 0 | 19.3
| 54.33 | 95.08 | 0
2097152 | 2097152 | uint8 | sum | -1 | 24.38 | 86.01 | 150.51 | 0 |
24.52 | 85.53 | 149.67 | 0
4194304 | 4194304 | uint8 | sum | -1 | 37.52 | 111.78 | 195.61 | 0 |
37.76 | 111.08 | 194.39 | 0
8388608 | 8388608 | uint8 | sum | -1 | 62.4 | 134.44 | 235.26 | 0 |
62.56 | 134.1 | 234.67 | 0
16777216 | 16777216 | uint8 | sum | -1 | 110.2 | 152.22 | 266.39 | 0 |
110.3 | 152.04 | 266.08 | 0
33554432 | 33554432 | uint8 | sum | -1 | 199.8 | 167.94 | 293.9 | 0 |
197.5 | 169.88 | 297.29 | 0
67108864 | 67108864 | uint8 | sum | -1 | 386.3 | 173.73 | 304.03 | 0 |
378.4 | 177.37 | 310.39 | 0
134217728 | 134217728 | uint8 | sum | -1 | 758 | 177.07 | 309.87 | 0 |
741.1 | 181.12 | 316.95 | 0
268435456 | 268435456 | uint8 | sum | -1 | 1500.1 | 178.95 | 313.16 | 0
| 1466.2 | 183.09 | 320.4 | 0
536870912 | 536870912 | uint8 | sum | -1 | 2991.7 | 179.45 | 314.04 | 0
| 2924.8 | 183.56 | 321.23 | 0

---------

Co-authored-by: Qinghua Zhou <qinghuahzhou@microsoft.com>
 2026-02-13 10:49:25 -08:00
Binyang Li
bd68319e3e Refactor algo selection logic and introduce symmetric_memory env (#741) 
This PR refactors the algorithm selection logic in MSCCL++ and
introduces support for symmetric memory configuration through
environment variables.


1. Algorithm Selection Refactoring
Use separate class for algo selection. Could introduce more complex
logic for algo selection based on message size, arch, if cuda graph is
enabled and memory allocation method

2. Symmetric Memory Support
Introduced symmetricMemory parameter in algorithm context key
generation. Remove disableChannelCache env as is ambiguous

3. Add new args for build_default_algorithms 
Add flag_buffer, and flag_buffer_size args to build default algorithm.
Then we could use unified flag buffer for different algorithms, avoid
application hanging when switch algo for different message size.

---------

Co-authored-by: chhwang <8018170+chhwang@users.noreply.github.com>
Co-authored-by: Qinghua Zhou <qinghuazhou@microsoft.com>
Co-authored-by: Caio Rocha <caiorocha@microsoft.com>
 2026-02-12 19:06:18 -08:00