Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
84 commits
Select commit Hold shift + click to select a range
4c11786
ggml-cuda: add internal AllReduce provider for tensor parallelism
scutler-nv Apr 22, 2026
52af99d
llama-bench: add --allreduce flag to select AllReduce provider
scutler-nv Apr 22, 2026
8c0a791
llama-bench: rename --allreduce to --reduction-provider / -rp
scutler-nv Apr 22, 2026
10c4757
llama-bench: pass WARN/ERROR log messages through in non-verbose mode
scutler-nv Apr 22, 2026
2c1a1db
cmake: improve NCCL detection for source-tree builds, add static/dyna…
scutler-nv Apr 22, 2026
c5207a4
ggml-cuda: add AllReduce hang watchdog (GGML_CUDA_AR_WATCHDOG)
scutler-nv Apr 22, 2026
172cba2
ggml-cuda: fix intermittent AllReduce hang on Blackwell PCIe
scutler-nv Apr 22, 2026
3b584e1
ggml-cuda: fix watchdog shutdown ordering and pipeline_free drain
scutler-nv Apr 23, 2026
860ee2a
ggml-cuda: replace event-based watchdog with per-GPU ring buffer
scutler-nv Apr 23, 2026
433c35a
fix: normalize line endings to LF (undo Windows CRLF conversion)
scutler-nv Apr 23, 2026
bc8b080
.gitattributes: force LF line endings to prevent Windows CRLF conversion
scutler-nv Apr 23, 2026
8da7e74
ggml-cuda: move GGML_CUDA_AR_WATCHDOG from CMake option to local define
scutler-nv Apr 23, 2026
372d408
unify kernel debug paths
scutler-nv Apr 23, 2026
5028250
use __threadfence_system explicitly (not in ggml_cuda_ar_signal_set)
scutler-nv Apr 23, 2026
cfdb0a6
preferentially use internal reduction for <=2 GPUs
scutler-nv Apr 23, 2026
239f20e
templatize the main kernel to support fp16/bf16
scutler-nv Apr 23, 2026
cfcaee3
restore llama-bench.cpp changes
scutler-nv Apr 23, 2026
bb01919
revert CMakeLists changes
scutler-nv Apr 23, 2026
ebc31bc
remove notes from repo
scutler-nv Apr 23, 2026
014ad9f
remove dead warmup code
scutler-nv Apr 23, 2026
a6981b5
fix comments
scutler-nv Apr 23, 2026
77c0e36
improve reduction provider fallback code
scutler-nv Apr 23, 2026
43743e5
add messages for allreduce fallback
scutler-nv Apr 23, 2026
2573b7b
rework reduction provider init to not call ncclCommInitAll if using t…
scutler-nv Apr 24, 2026
4d7736c
fix case where a given tensor has not been computed
scutler-nv Apr 24, 2026
f046004
add chunked mode to the kernel for unlimited vector size
scutler-nv Apr 24, 2026
fbcae51
rework a few checks/fallbacks
scutler-nv Apr 24, 2026
f3d321f
Merge branch 'master' into dev/internal-allreduce
scutler-nv Apr 24, 2026
7449c1d
various small cleanups
scutler-nv Apr 24, 2026
2b91e21
allow disabling CUDA reductions completely (falling back to the non-C…
scutler-nv Apr 24, 2026
485ee23
simplify reduction provider selection
scutler-nv Apr 24, 2026
892b2e3
minor simplifications
scutler-nv Apr 24, 2026
b34adf4
more cleanups/fixes
scutler-nv Apr 24, 2026
b8caad3
prototype alternate path for large reductions
scutler-nv Apr 25, 2026
b0bdf83
chunked version of large reduction path
scutler-nv Apr 25, 2026
77b96e6
use bf16 for large reductions
scutler-nv Apr 27, 2026
0805e6a
experimental reduction using cudaMemcpyPeerAsync (slightly slower)
scutler-nv Apr 28, 2026
97eae7c
revert experimental change
scutler-nv Apr 28, 2026
3e8b263
add combined conversion/reduction kernel
scutler-nv Apr 28, 2026
80faf56
add bf16 wire format for single kernel mode
scutler-nv Apr 28, 2026
210165b
experimental on-stream small reduction kernel
scutler-nv Apr 28, 2026
fbc2bd7
double buffer arrival slots, use token (incrementing) method
scutler-nv Apr 28, 2026
3a41c7b
double buffer host_buf for small reductions
scutler-nv Apr 28, 2026
330c14c
put in waits for use of host_mem in large reduction case (prevents st…
scutler-nv Apr 28, 2026
2b91d02
remove watchdog code
scutler-nv Apr 28, 2026
ff4ed48
various cleanups / dead code removal
scutler-nv Apr 28, 2026
25ff620
fix fp16 mode
scutler-nv Apr 28, 2026
26c3d01
fix some comments/logging statements
scutler-nv Apr 28, 2026
83b9bac
use increasing token scheme for arrival signals
scutler-nv Apr 28, 2026
1fe1528
add top-level comment to allreduce.cu
scutler-nv Apr 28, 2026
630a800
improve top-level comment in allreduce.cu
scutler-nv Apr 28, 2026
6d4b998
fix comments in ggml_cuda_ar_kernel
scutler-nv Apr 28, 2026
63608b7
improve event handling for hostmem buffer usage tracking
scutler-nv Apr 29, 2026
be783f9
change ev_pool to fixed 2D array
scutler-nv Apr 29, 2026
f4a4ce2
Merge branch 'master' into dev/internal-allreduce
scutler-nv Apr 29, 2026
56a87ee
add chunked memcpy fallback for extra-large reductions (>32 MB)
scutler-nv Apr 30, 2026
59c51b3
change thresholds for copy-engine path and bf16 demotion
scutler-nv Apr 30, 2026
28a71d3
multi-block kernel test
scutler-nv Apr 30, 2026
b35a093
more fine-tuning for chukn-size, etc.
scutler-nv Apr 30, 2026
5d3df69
various fixes for PR review
scutler-nv Apr 30, 2026
bfa0b69
more PR fixes
scutler-nv May 1, 2026
0f9bea3
fix semantics of all host mappings
scutler-nv May 1, 2026
132934e
require ampere+
scutler-nv May 1, 2026
1854ebc
small cleanups
scutler-nv May 1, 2026
6e9319c
properly use host pointer for src/dst in cudaMemcpy calls
scutler-nv May 4, 2026
a19d711
allreduce: lazy-init the internal pipeline on first use
scutler-nv May 5, 2026
5d1b22a
allreduce: assert n_backends == 2 instead of soft-fallback
scutler-nv May 5, 2026
727b141
rework reduction provider selection. internal/nccl is OS dependent; m…
scutler-nv May 6, 2026
e85a280
remove unneeded Turing arch check (llama.cpp doesn't even compile pre…
scutler-nv May 6, 2026
cb3e521
allreduce: ASCII-only comments and ggml_cuda_cast for value conversions
scutler-nv May 6, 2026
d11b178
allreduce: use ggml_cuda_memcpy_1 for the chunked-kernel vector copies
scutler-nv May 6, 2026
7db79ff
allreduce: assert cuda_ctx->device matches the pipeline's device
scutler-nv May 6, 2026
605f5e0
allreduce: expand one-liner for loops to braced bodies
scutler-nv May 7, 2026
77e9b14
allreduce: rename template parameters Tdst/Twire/Tsrc -> T_dst/T_wire…
scutler-nv May 7, 2026
b5e18bc
allreduce: drop hyphen in 'chunked-kernel' across comments
scutler-nv May 7, 2026
f4c8b26
allreduce: use ggml_cuda_get_max_cpy_bytes() instead of hardcoded 16
scutler-nv May 7, 2026
d816e47
Merge remote-tracking branch 'origin/master' into dev/internal-allreduce
scutler-nv May 7, 2026
29b11cf
ggml-cuda: PR review fixes -- annotate #endif, fix stale comment, ass…
scutler-nv May 8, 2026
6ae3093
simplify the init logic
scutler-nv May 8, 2026
d64b3b1
address some other PR requests
scutler-nv May 8, 2026
2dd828a
ggml-cuda: stub internal AllReduce on HIP/MUSA, drop pre-Ampere menti…
scutler-nv May 9, 2026
ced4924
allreduce: guard __nanosleep on Volta+ and reject pre-Volta devices a…
scutler-nv May 9, 2026
db55160
allreduce: fix CI -Werror warnings (sign-compare, format, restrict al…
scutler-nv May 9, 2026
8fd5f57
ggml-cuda: drop unused-function warning by guarding try_allreduce_ncc…
scutler-nv May 9, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Loading
Loading