9f654f6cf5
graph: restrict MAXCHANNELS for certain platforms
2024-06-03 13:05:01 -05:00
0948eecbba
Changing channel stride for MI300X multinode ( #1196 )
...
* Shuffling MI300X multi-node channels
* Updating tree channel logic
2024-06-03 10:00:55 -06:00
354e0b29a6
Addressing possible out-of-bounds mem access during channel duplication ( #1193 )
2024-05-30 14:02:14 -06:00
73221b4230
Add ring simple chunk size tuning ( #1180 )
...
* Add ring simple chunk size tuning
* modifying the tuning table to improve the performance of broadcast for 8MB to 32MB for single-node MI300X after ring simple chunk size tuning
* modifying the tuning table to improve the performance of reduce for 1MB to 4MB for single-node MI300X after ring simple chunk size tuning
---------
Co-authored-by: PedramAlizadeh <pmohamma@amd.com >
2024-05-29 07:59:47 -07:00
73acf3eeec
modifying the tuning table to improve the performance of broadcast for 1MB to 64MB for single-node MI300X ( #1172 )
2024-05-08 15:49:33 -04:00
408278209d
Adding ASAN changes to address memory leak issue" ( #1170 )
...
Co-authored-by: akolliasAMD <akollias@amd.com >
2024-05-08 09:16:00 -05:00
b18784d8b8
Add compiler warning for uninitialized variable and fix ( #1163 )
...
* Add compiler warning for uninitialized variable and fix
* Add -Wsometimes-uninitialized
* Convert warning to error
2024-05-08 07:00:25 -07:00
f679db6ff6
Use normal permute path when one NIC per GPU ( #1171 )
2024-05-08 06:59:57 -07:00
b513c3970a
Bypass NVIDIA Ampere related tuning ( #1165 )
2024-05-03 17:57:16 -07:00
bb58b1c258
Fix ignore NUMA not being observed for NICs during model matching ( #1164 )
2024-05-03 16:42:07 -07:00
9e0c9b4ed8
Replace __HIP_PLATFORM_HCC__ with __HIP_PLATFORM_AMD__ ( #1154 )
2024-04-25 07:19:18 -07:00
e1a835910e
Merge remote-tracking branch 'nccl/master' into develop
2024-04-23 13:34:00 -07:00
220066197a
Use hipExtMallocWithFlags to allocate host memory on APU ( #1149 )
...
Also use SM60 as CUDA compatibility level.
2024-04-17 16:56:38 -07:00
4cb62f999a
Rail optimization for rings ( #1140 )
...
- Modifies the ring creation algorithm to be friendlier to rail-optimized topologies (should not affect classic fabric topologies)
2024-04-15 12:03:57 -06:00
df98a6957d
Add another Rome model ( #1095 )
2024-02-28 10:46:05 -08:00
48bb7fec79
2.20.5-1
...
Fix UDS connection failure when using ncclCommSplit. Issue #1185
2024-02-26 02:52:39 -08:00
74f9e5db64
Add new GPU model ( #1080 )
2024-02-23 12:19:42 -08:00
2fb12a9358
Merge pull request #1079 from BertanDogancay/2.19.4-sync
...
2.19.4 Sync
2024-02-16 09:50:11 -07:00
bac57421c7
Allow bus id to be null ( #1085 )
...
* Allow bus id to be null
2024-02-15 16:36:51 -07:00
b6475625fb
2.20.3-1
...
Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.
2024-02-13 04:22:38 -08:00
d999d9ad21
Merge remote-tracking branch 'rccl/develop' into 2.19.4
2024-02-09 11:31:03 -06:00
5669b0d7b6
2.18.5 fix ( #1077 )
...
* Revert "Revert "2.18.5-1""
This reverts commit 767fde8210
2024-02-09 09:18:38 -08:00
704c9ef0d1
Doubling P2P channels per peer on single node gfx94x only ( #1074 )
2024-02-07 14:05:57 -08:00
1d989f6524
Doubling P2P channels per peer on single node only ( #1069 )
2024-02-02 12:41:00 -08:00
1a134b283b
Merge remote-tracking branch 'rccl/develop' into 2.19.4
2024-01-31 11:53:10 -06:00
9ff53eeeae
Merge remote-tracking branch 'nccl/master' into develop
2024-01-30 14:43:43 -08:00
ccfb35fa6d
modifying the tuning table to improve the performance of allreduce for 8MB and 16MB for single-node MI300X ( #1063 )
2024-01-26 09:05:53 -05:00
ffde530af5
Increase P2P channels per peer ( #1060 )
2024-01-25 11:21:58 -08:00
81ddf9de89
Merge remote-tracking branch 'nccl/v2.19' into develop
2024-01-24 15:25:33 -08:00
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
735178c1fe
cover more gpu/nic mapping cases ( #1037 )
2024-01-10 08:01:37 -08:00
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
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
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
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
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
dbcba2923b
Use parallel init of LDS and adjust P2P channels for gfx94x ( #943 )
...
* Use parallel init of LDS and adjust P2P channels for gfx94x
* Move another init to parallel
* Fix NCCL_NCHANNELS_PER_PEER setting
2023-11-03 16:06:49 -07:00
1e5ca6820b
Fix gcnArchName bug in topology dump ( #937 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
2023-10-28 12:30:36 -05:00
f22df90e5c
remove gcnArch support ( #920 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
2023-10-26 12:09:15 -05:00
c4e65fd382
Add missing gfx942 support ( #927 )
2023-10-23 12:04:37 -07:00
4278a9918b
Update rome models ( #922 )
2023-10-18 17:28:01 -07:00
a6ff4618c7
Revert "Remove 2H4P condition from P2P channels adjustment ( #890 )" ( #904 )
...
This reverts commit 16dd05a58a
2023-10-04 09:46:11 -06:00
Nusrat Islam
gilbertlee-amd
Wenkai Du
Pedram Alizadeh
mberenjk
BertanDogancay
Sylvain Jeaugey
Bertan Dogancay
akolliasAMD
Wenkai Du
Hossein Pourreza
Ziyue Yang
pradeep-ramanna
Alexander Grund
Nilesh M Negi