7987015a19
Revert "msccl: build same number of kernels as in ROCm 5.7" ( #1058 )
...
This reverts commit f960174d03be7e5174baa83b256526d388a38842.
2024-01-24 08:43:50 -08:00
5564d65e71
Use binary search for direct function calls ( #1057 )
...
* Use binary search for direct function calls
* fix scratch mem issue on MI300
2024-01-22 17:37:56 -07:00
c4dbf8a914
Fix collective trace when rccl is configured ( #1056 )
...
* Fix collective trace when rccl is configured
2024-01-22 09:26:44 -07:00
7e25d5bc55
Use new HIP graph API compatible with CUDA 11030 ( #991 )
...
* Use new HIP graph API compatible with CUDA 11030
* Update dependency to ROCm 6.1
* Fix single stream use case
2024-01-21 19:00:50 -08:00
8b97a20943
COLLECTIVES: Switch to unroll 2 for MI300 ( #1051 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
2024-01-19 12:16:05 -06:00
28d9b170c9
[DEV] Configure functions in RCCL ( #986 )
...
* configure functions in rccl
2024-01-18 15:07:16 -07:00
3325f96c56
Only use full MAXCHANNELS for gfx94x ( #1050 )
2024-01-17 09:00:49 -08:00
b08124c85d
adding rccl tuning parameters for MI300X gfx942 with 8 GPUs single and multi-node ( #1047 )
2024-01-16 13:44:32 -05:00
261707d90a
Add option to force enable network transport on single node ( #1046 )
2024-01-16 07:54:18 -08:00
767fde8210
Revert "2.18.5-1"
...
This reverts commit 559b70f86c
2024-01-12 16:54:19 +00:00
cf248d9402
Addressing the compiler warning ( #988 )
2024-01-10 14:59:40 -07:00
735178c1fe
cover more gpu/nic mapping cases ( #1037 )
2024-01-10 08:01:37 -08:00
5851ae5974
Re-enable L128 on gfx90a of compiler supports it ( #1036 )
2024-01-10 08:01:11 -08:00
249e9f7f65
Un-escaped character causes error with address sanitizer builds ( #992 )
...
Signed-off-by: Nilesh M Negi <Nilesh.Negi@amd.com >
Co-authored-by: Jenkins <jenkins-compute@amd.com >
2024-01-09 13:28:32 -06:00
aa5c84c997
Merge pull request #1022 from PedramAlizadeh/sync_nccl_2.18.6
...
Sync to nccl 2.18.6
2024-01-09 13:29:29 -05:00
d9871d171b
msccl: use custom reduce function ( #1033 )
2024-01-08 14:53:12 -08:00
f7e39fced2
Doubling buffer size to fix NCCL INFO corruption with increased channels ( #1035 )
2024-01-08 08:14:33 -08:00
e5bf56c6d8
Increase stack size for gfx906 ( #1034 )
...
Occationally "Memory access fault by GPU node-8 (Agent handle: 0x23a5640) on address 0x7f461ec00000. Reason: Page not present or supervisor privilege" can be seen from gfx906 CI
2024-01-07 20:25:02 -08:00
70bbeb4773
Fix MSCCL multi-node ( #1032 )
...
1) Move needsProxy initialization before mscclSetupConnections since the latter
will revise it later.
2) Remove mscclAvailable check in net.cc since it's no more required and caused
non-shared buffer allocated for MSCCL which is not expected.
2024-01-05 17:03:43 -08:00
abf265a911
Rework barriers and adjust scope of atomics ( #1019 )
2024-01-04 08:18:48 -08:00
0a53077c9c
Improve MSCCL algorithms ( #1023 )
2024-01-03 14:51:34 -08:00
f4858e14b2
rearranged how the min and max functions are part of msccl ( #1025 )
...
* rearranged how the min and max functions are part of msccl
* added more coverage on in place graph tests
2023-12-21 08:58:33 -07:00
0d515f9388
resolved conflicts, fixed the localNetCount/0 bug
2023-12-18 08:11:34 +00:00
655742a3a6
Fully disable MSCCL when machine is not matched ( #1017 )
...
* Disable MSCCL algorithm meta loading when machine is not matched
* fully disable init
* fix potential segfault
2023-12-13 08:36:21 -08:00
53d807a5b9
msccl: disable on multi-node ( #1018 )
2023-12-13 07:41:40 -08:00
81602814a7
msccl: fix data corruption with MTYPE_RW ( #1014 )
2023-12-11 20:33:15 -08:00
7965c8b53c
Fix memory fence and use non-temporal store ( #1007 )
...
* Fix memory fence and use non-temporal store
* Use amdgcn builtin instead of inline asm
* Move threadfence location
* Revert changes to gfx90a
* Rework gfx90a change
* Apply changes to gfx94x
2023-12-09 12:16:08 -08:00
c002f20029
Fix MSCCL scratch allocation ( #1010 )
2023-12-08 17:47:10 -06:00
baadda4bd8
Relax workgroup barrier implementation for MSCCL send/recv ops. ( #997 )
...
* Trim logic.
* Revert "Trim logic."
This reverts commit 8f2dba6c764108acf2bf5428366b9f41d4d206b9.
* Introduce MSCCL template parameters to send / recv.
* Address review feedbacks.
2023-12-08 17:46:53 -06:00
12c08fc52a
msccl: build same number of kernels as in ROCm 5.7 ( #1005 )
...
Removed fullOps kernels from build
2023-12-07 13:36:04 -06:00
293f0fb752
Use a map to host scratch buffers ( #1004 )
...
* Use a map to host scratch buffers
* Address review feedbacks. Deliberately keep mscclSetupScratch function.
2023-12-05 13:15:28 -06:00
bc44e3faa7
Fix gcnArch bug in IFC mix build ( #998 ) ( #1002 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
2023-12-04 16:20:22 -06:00
7c0f49a878
IFC mix build ( #998 )
2023-12-02 18:49:52 -07:00
4ba65d1d6a
Increase max channles to 64 ( #993 )
2023-12-01 16:01:11 -08:00
0b53f79196
Fix GPU to NIC mapping for peertopeer ( #994 )
2023-12-01 08:00:17 -08:00
e44e112a17
Fix mscclAlgoHandle not initialized issue ( #995 )
2023-12-01 07:58:01 -08:00
4bb0b4a380
Move MSCCL algorithm loading to initialization to workaround HIP graph conflict ( #982 )
...
* MSCCL: pre-specify channels and pre-load algorithms
* add mutex
* fix bug
* clean include
* disable all-gathers temporarily
2023-11-30 09:47:20 -08:00
56ce9ef05f
recreated pr 914 to work with current develop branch ( #979 )
2023-11-28 16:33:47 -07:00
50b2dd9fd7
Add special handling of gfx940 ( #976 )
...
* Add special handling of gfx940
* Update ring base
2023-11-22 15:07:36 -08:00
b6d7438d31
Merge remote-tracking branch 'origin/master'
2023-11-20 05:07:23 -08:00
569d3f7d59
msccl: allocate scratch as ext-scope fine-grained ( #968 )
2023-11-16 09:57:25 -06:00
bc8661f092
Fix kernel command line warnings ( #961 )
...
* Fix kernel command line warnings
* Remove while loop
2023-11-15 18:01:12 -08:00
7fc891bc8d
Fix MSCCL work FIFO allocation with HIP graph enabled ( #967 )
2023-11-15 16:43:28 -08:00
198f14923b
Check to support older ROCm versions ( #963 )
2023-11-15 12:36:31 -07:00
7ae95db5b8
Optimize MSCCL all-gather algorithms for gfx942 ( #964 )
2023-11-15 08:18:59 -08:00
df128879a6
Optimize MSCCL reduce primitive switching for gfx942 ( #962 )
...
* Optimize reduce primitive switching for gfx942
* address comment
2023-11-15 08:18:44 -08:00
cece6415b0
Fix use of CPUID overwriting registers in use.
...
CPUID writes to EAX, EBX, ECX, and EDX so the inline-asm must state that.
Otherwise currently in-use register might get overwritten which may
cause all kinds of failures like segfaults or wrong results.
Alternatively `__cpuid` can be used which avoids this and related issues.
So do that as suggested in the GCC issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112513
2023-11-14 12:38:02 +01:00
88d44d777f
2.19.4-1
...
Split transport connect phase into multiple steps to avoid port
exhaustion when connecting alltoall at large scale. Defaults to 128
peers per round.
Fix memory leaks on CUDA graph capture.
Fix alltoallv crash on self-sendrecv.
Make topology detection more deterministic when PCI speeds are not
available (fix issue #1020 ).
Properly close shared memory in NVLS resources.
Revert proxy detach after 5 seconds.
Add option to print progress during transport connect.
Add option to set NCCL_DEBUG to INFO on first WARN.
2023-11-13 10:36:12 -08:00
5a800e00cd
msccl: enable basic collective trace ( #959 )
...
To avoid increasing number of kernels, colltrace is only enabled with
RCCL_MSCCL_FORCE_FULLOPS=1
2023-11-08 20:14:28 -08:00
efc42d9045
Use send instead of sendWithBarrier. ( #727 )
2023-11-07 13:47:24 -06:00
Wenkai Du
Bertan Dogancay
Nilesh M Negi
Pedram Alizadeh
Hossein Pourreza
Ziyue Yang
akolliasAMD
Wen-Heng (Jack) Chung
pradeep-ramanna
Sylvain Jeaugey
Alexander Grund