AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

NCCL 2.26.2-1

Browse Source 
Profiler improvements
 * Add events for CUDA kernel start and end.
 * Allow network plugins to generate profiling events
 * Enable profiling on a per-operation basis, rather than per-communicator.
 * Add support for graph capturing.

Add implicit launch order
 * Allow to prevent deadlocks when using multiple NCCL communicators per
   device by implicitly ordering NCCL operations using the host program
   order. Disabled by default, set NCCL_LAUNCH_ORDER_IMPLICIT=1 to enable.
 * Add a complementary mechanism to detect host threads racing to launch
   to the same device. Enabled by default, set NCCL_LAUNCH_RACE_FATAL=0 to
   disable.

Optimize the PAT algorithm
 * Separate the computation and execution of PAT steps on different warps,
   allowing to run up to 16 PAT steps in parallel to significantly
   accelerate PAT and reduce its linear part.

Add support for setting QoS per communicator
 * Add a new trafficClass field to the communicator configuration, to
   allow the application to select a particular traffic class for a
   given communicator. The meaning of the traffic class is
   network-specific and should be set in accordance with the network
   configuration.
 * For the IB/RoCE plugin, existing config variables such as NCCL_IB_SL
   and NCCL_IB_TC take precedence.

Allow to enable GPU Direct RDMA specifically on C2C platforms
 * Disabled by default, set NCCL_NET_GDR_C2C=1 to enable.

Do not disable user buffer registration unless PXN is really used
 * Only disable UB when a communicator has more than one rank per
   node on any node.

RAS subsystem improvements
 * Report operation counts separately for each collective operation type.
 * Provide details about missing communicator ranks and reliably
   distinguish ranks that are no longer a given communicator's members
   (now reported as NOCOMM) from those that failed to respond.

Add support for timestamps to NCCL diagnostic messages
 * On by default for WARN messages; NCCL_DEBUG_TIMESTAMP_LEVELS can be
   used to enable them for other debug levels as well.
 * The format can be changed using the NCCL_DEBUG_TIMESTAMP_FORMAT config
   variable.

Reduce the memory usage with NVLink SHARP (NVLS)
 * Potentially save hundreds of MBs of device memory, considering the
   multicast buffer size granularity separately from the address alignment.

Update performance tuning for recent Intel CPUs
 * Improve algorithm/protocol selection on recent CPUs such as Emerald
   Rapids and Sapphire Rapids.

Improve channel scheduling when mixing LL and Simple operations.
 * Make LL operations account for 4x more traffic to ensure LL and simple
   operations complete at the same time.

Refactor the plugin code
 * Clean up and harmonize the support code across the network, tuner,
   and profiler plugins.

Add support for comment lines (starting with #) in the nccl.conf file
* Issue #1540.

Make user buffer registration problems print an INFO instead of a WARN.

Drop support for network plugin interface version 5.

Fix a race condition with split-shared communicators
 * NCCL could hang during connection setup if multiple communicators
   were grouped together that share resources.

Fix a performance regression when using NCCL_CROSS_NIC=1
 * NCCL would unnecessarily alternate rings, breaking the GPU-NIC
   associations.

Make GID index detection code more resilient
 * Dynamic GID detection code was giving up too soon if the
   detected index was not available (e.g., wasn't mapped to the
   container's sysfs).
 * Issues #1538, #1573.

Fix a race condition with non-blocking operation
 * Fix issue when creating a non-blocking communicator after a non-
   blocking collective operation on another communicator.

Fix shared memory usage on recent Blackwell GPUs.
 * Issues NVIDIA/nccl-tests#287, NVIDIA/nccl-tests#291, #1637.

Fix an error with NIC fusion and IB SHARP when recreating communicators
 * Disable the unloading of network plugins

Make the auto-merge failures in the NIC fusion non-fatal
 * This could happen when trying to merge IB and RoCE devices.

Fixes to ncclCommAbort
 * Fix hangs due to the progress thread spinning indefinitely on the
   network progress.
 * Reduce the abort time by up to two orders of magnitude.

Fix a crash when libnccl.so was dynamically unloaded
 * The RAS subsystem was missing a clean-up handler.

Fix a hang if the network plugin's test() call returns an error.

Fix a hang on heterogeneous architectures
 * Ensure we harmonize the tuning to avoid different tuning choices,
   causing a hang.

Fix double-free on failed ncclCommInitRank and ncclCommFinalize.

Fix a potential list traversal bug during a group launch of multiple
communicators
 * Issue #1599.

Unify the handling of NCCL configuration variables
 * Under rare circumstances, some variables specified in the config file
   could be ignored.


[ROCm/rccl commit: f44ac759fe]
This commit is contained in:
Kamil Iskra
2025-03-12 13:46:21 -07:00
parent 70bd1af305
commit 59f0bf79c5
116 changed files with 7522 additions and 5278 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/strongstream.h
+46 -52
View File
@@ -10,13 +10,24 @@
#include "nccl.h"
#include "checks.h"
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdint.h>
// ncclCudaContext: wraps a CUDA context with per-context state.
struct ncclCudaContext;
// Get a ncclCudaContext to track the currently active CUDA context.
ncclResult_t ncclCudaContextTrack(struct ncclCudaContext** out);
// Drop reference.
void ncclCudaContextDrop(struct ncclCudaContext* cxt);
/* ncclCudaGraph: Wraps a cudaGraph_t so that we can support pre-graph CUDA runtimes
* easily.
*/
struct ncclCudaGraph {
#if CUDART_VERSION >= 11030
cudaStream_t origin;
cudaGraph_t graph;
unsigned long long graphId;
#endif
@@ -25,6 +36,7 @@ struct ncclCudaGraph {
inline struct ncclCudaGraph ncclCudaGraphNone() {
struct ncclCudaGraph tmp;
#if CUDART_VERSION >= 11030
tmp.origin = nullptr;
tmp.graph = nullptr;
tmp.graphId = ULLONG_MAX;
#endif
@@ -33,7 +45,7 @@ inline struct ncclCudaGraph ncclCudaGraphNone() {
inline bool ncclCudaGraphValid(struct ncclCudaGraph graph) {
#if CUDART_VERSION >= 11030
return graph.graph != nullptr;
return graph.graphId != ULLONG_MAX;
#else
return false;
#endif
@@ -57,60 +69,37 @@ ncclResult_t ncclCudaGraphAddDestructor(struct ncclCudaGraph graph, cudaHostFn_t
* streams unfit for the use of serializing access to a persistent resource.
* Strong streams have been introduced to address this need.
*
* - All updates to a strong stream must be enclosed by a Acquire/Release pair.
* All updates to a strong stream must be enclosed by a Acquire/Release pair.
*
* - The Acquire, Release, and all updates take a ncclCudaGraph parameter
* indicating the currently capturing graph (or none). This parameter must be
* the same for the entire sequence of {Acquire; ...; Release}.
* Acquire retrieves a "work" stream (cudaStream_t) which may be used to add
* work.
*
* - An {Acquire; ...; Release} sequence must not be concurrent with any
* other operations against the strong stream including graph launches which
* reference this stream.
* Release publishes the work streams work into the strong stream. The Release
* must be issued by the same thread that did the Acquire.
*/
struct ncclStrongStream;
ncclResult_t ncclStrongStreamConstruct(struct ncclStrongStream* ss);
ncclResult_t ncclStrongStreamDestruct(struct ncclStrongStream* ss);
// Acquire-fence the strong stream.
// Acquire the strong stream. Upon return `*workStream` will be usable to add work.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamAcquire(
struct ncclCudaGraph graph, struct ncclStrongStream* ss
struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent, cudaStream_t* workStream
);
// Acquire-fence the strong stream assuming no graph is capturing. This permits
// the caller to enqueue directly to the `ss->cudaStream` member using native CUDA
// calls. Strong stream still must be released via:
// ncclStrongStreamRelease(ncclCudaGraphNone(), ss);
ncclResult_t ncclStrongStreamAcquireUncaptured(struct ncclStrongStream* ss);
// Release-fence of the strong stream.
ncclResult_t ncclStrongStreamRelease(struct ncclCudaGraph graph, struct ncclStrongStream* ss);
// Add a host launch to the stream.
ncclResult_t ncclStrongStreamLaunchHost(
struct ncclCudaGraph graph, struct ncclStrongStream* ss,
cudaHostFn_t fn, void* arg
);
// Add a kernel launch to the stream.
ncclResult_t ncclStrongStreamLaunchKernel(
struct ncclCudaGraph graph, struct ncclStrongStream* ss,
void* fn, dim3 grid, dim3 block, void** args, size_t sharedMemBytes
// Get the workStream for an already acquired strong stream.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamAcquiredWorkStream(
struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent, cudaStream_t* workStream
);
// Cause `a` to wait for the current state `b`. Both `a` and `b` must be acquired.
// `b_subsumes_a` indicates that all work in `a` is already present in `b`, thus
// we want to fast-forward `a` to be a clone of `b`. Knowing this permits the
// implementation to induce few graph dependencies.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, struct ncclStrongStream* a, struct ncclStrongStream* b, bool b_subsumes_a=false
);
// `b` must be capturing within `graph`.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, struct ncclStrongStream* a, cudaStream_t b, bool b_subsumes_a=false
);
// `a` must be capturing within `graph`.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, cudaStream_t a, struct ncclStrongStream* b, bool b_subsumes_a=false
// Release of the strong stream.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamRelease(struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent);
ncclResult_t ncclStreamWaitStream(
cudaStream_t a, cudaStream_t b, cudaEvent_t scratchEvent
);
// Synchrnoization does not need the strong stream to be acquired.
@@ -118,23 +107,28 @@ ncclResult_t ncclStrongStreamSynchronize(struct ncclStrongStream* ss);
////////////////////////////////////////////////////////////////////////////////
struct ncclStrongStreamGraph; // internal to ncclStrongStream
struct ncclStrongStreamCapture; // internal to ncclStrongStream
struct ncclStrongStream {
// Used when not graph capturing.
cudaStream_t cudaStream;
// The stream to use for non-captured work.
cudaStream_t liveStream;
void* liveAcquiredBy;
#if CUDART_VERSION >= 11030
// This stream ever appeared in a graph capture.
bool everCaptured;
pthread_mutex_t lock;
struct ncclStrongStreamCapture* captureHead;
// The event used to establish order between graphs and streams. During acquire
// this event is waited on, during release it is recorded to.
cudaEvent_t serialEvent;
// This stream ever appeared in a graph capture.
bool everCaptured;
// Tracks whether serialEvent needs to be recorded to upon Release().
bool serialEventNeedsRecord;
struct ncclStrongStreamGraph* graphHead;
#else
cudaEvent_t scratchEvent;
#endif
};
struct ncclCudaContext {
struct ncclCudaContext* next;
CUcontext hcontext;
int refCount;
struct ncclStrongStream launchOrder;
};
#endif