Question about the ENABLE_INTRA_NODE_COMM for speculative decoding about gpt-fast

[jianc99]

And here is the env version @yifuwang :
PyTorch version: 2.5.0.dev20240613+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: version 3.27.7
Libc version: glibc-2.31

Python version: 3.11.9 | packaged by conda-forge | (main, Apr 19 2024, 18:36:13) [GCC 12.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-1048-aws-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 535.104.12
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
Stepping: 7
CPU MHz: 2999.998
BogoMIPS: 5999.99
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 1.5 MiB
L1i cache: 1.5 MiB
L2 cache: 48 MiB
L3 cache: 71.5 MiB
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Vulnerability Gather data sampling: Unknown: Dependent on hypervisor status
Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines, STIBP disabled, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-triton==3.0.0+45fff310c8
[pip3] torch==2.5.0.dev20240613+cu121
[pip3] torchaudio==2.4.0.dev20240617+cu121
[pip3] torchvision==0.19.0.dev20240617+cu121
[conda] numpy 1.26.4 pypi_0 pypi
[conda] pytorch-triton 3.0.0+45fff310c8 pypi_0 pypi
[conda] torch 2.5.0.dev20240613+cu121 pypi_0 pypi
[conda] torchaudio 2.4.0.dev20240617+cu121 pypi_0 pypi
[conda] torchvision 0.19.0.dev20240617+cu121 pypi_0 pypi

[jianc99]

Yes, in my case, it always hung at a specific iteration as well. And actually I also encountered this problem for autoregressive decoding without speculation. But it doesn't occur each time, like a random event.

[yifuwang]

I believe pytorch/pytorch#129501 should fix this issue

[jianc99]

And this problem only occurs on 8xA100 and 4xA100, I tested on other machine like 2xA100 and 8xL40, the problem did't occur.

[yifuwang]

Hey @jianc99, I wasn't able to reproduce the issue on my setup. Can you post your GPU connectivity with nvidia-smi topo -m and post it here?

[jianc99]

Hi @yifuwang , here is the GPU connectivity info

	� GPU0	GPU1	GPU2	GPU3	CPU Affinity	NUMA Affinity	GPU NUMA ID�[0m
GPU0	 X 	NV12	NV12	NV12	0,48	0		N/A
GPU1	NV12	 X 	NV12	NV12	0,48	0		N/A
GPU2	NV12	NV12	 X 	NV12	0,48	0		N/A
GPU3	NV12	NV12	NV12	 X 	0,48	0		N/A

And the torch version I am using is 2.5.0.dev20240613+cu121. Whether or not compile doesn't affect the occur of the problem.

export MODEL_REPO=meta-llama/Llama-2-70b-hf
export DRAFT_MODEL_REPO=meta-llama/Llama-2-7b-hf
ENABLE_INTRA_NODE_COMM=1 torchrun --standalone --nproc_per_node=4 generate.py  --draft_checkpoint_path checkpoints/$DRAFT_MODEL_REPO/model.pth  --checkpoint_path checkpoints/$MODEL_REPO/model.pth --speculate_k 5 --prompt "def quicksort(arr):" --max_new_tokens 200 --num_samples 50 --temperature 0

The output is below, just running several iterations, then stuck, and all the GPU util keeps 100%.

def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 1: 19.81 sec total, 10.10 tokens/sec
Bandwidth achieved: 350.92 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 2: 12.96 sec total, 15.43 tokens/sec
Bandwidth achieved: 536.10 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 3: 13.00 sec total, 15.39 tokens/sec
Bandwidth achieved: 534.68 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 4: 12.59 sec total, 15.88 tokens/sec
Bandwidth achieved: 551.99 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 5: 12.64 sec total, 15.82 tokens/sec
Bandwidth achieved: 549.77 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 6: 12.95 sec total, 15.45 tokens/sec
Bandwidth achieved: 536.78 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 7: 12.73 sec total, 15.71 tokens/sec
Bandwidth achieved: 545.87 GB/s
def quicksort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = [x for x in arr[1:] if x < pivot]
    right = [x for x in arr[1:] if x >= pivot]
    return quicksort(left) + [pivot] + quicksort(right)


if __name__ == '__main__':
    arr = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
    print(quicksort(arr))
 package com.example.android.miwok;

import android.content.Context;
import android.media.AudioManager;
import android.media.MediaPlayer;
import android.os.Bundle;
import android.support.v4.app.Fragment;
import android
Time for inference 8: 12.73 sec total, 15.71 tokens/sec
Bandwidth achieved: 546.02 GB/s

[jianc99]

@yifuwang And just use the same code and remove ENABLE_INTRA_NODE_COMM=1, the code can successfully run.

[jianc99]

Add an error information. Sometimes there will be the error message like below. And sometimes it just stuck without any error message. Hope this will be helpful @yifuwang
2%|▏ | 4/200 [07:09<3:14:17, 59.48s/it][rank2]:[E622 01:14:05.084831011 ProcessGroupNCCL.cpp:607] [Rank 2] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=9, OpType=ALLREDUCE, NumelIn=1, NumelOut=1, Timeout(ms)=600000) ran for 600064 milliseconds before timing out.
[rank0]:[E622 01:14:05.084836161 ProcessGroupNCCL.cpp:607] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=9, OpType=ALLREDUCE, NumelIn=1, NumelOut=1, Timeout(ms)=600000) ran for 600064 milliseconds before timing out.
[rank3]:[E622 01:14:05.084837275 ProcessGroupNCCL.cpp:607] [Rank 3] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=9, OpType=ALLREDUCE, NumelIn=1, NumelOut=1, Timeout(ms)=600000) ran for 600065 milliseconds before timing out.
[rank0]:[E622 01:14:05.145697835 ProcessGroupNCCL.cpp:1664] [PG 0 (default_pg) Rank 0] Exception (either an error or timeout) detected by watchdog at work: 9, last enqueued NCCL work: 9, last completed NCCL work: 8.
[rank2]:[E622 01:14:05.145700867 ProcessGroupNCCL.cpp:1664] [PG 0 (default_pg) Rank 2] Exception (either an error or timeout) detected by watchdog at work: 9, last enqueued NCCL work: 9, last completed NCCL work: 8.
[rank3]:[E622 01:14:05.145704160 ProcessGroupNCCL.cpp:1664] [PG 0 (default_pg) Rank 3] Exception (either an error or timeout) detected by watchdog at work: 9, last enqueued NCCL work: 9, last completed NCCL work: 8.
[rank0]:[E622 01:26:52.726366069 ProcessGroupNCCL.cpp:1375] [PG 0 (default_pg) Rank 0] First PG on this rank that detected no heartbeat of its watchdog.
[rank2]:[E622 01:26:52.726387839 ProcessGroupNCCL.cpp:1375] [PG 0 (default_pg) Rank 2] First PG on this rank that detected no heartbeat of its watchdog.
[rank3]:[E622 01:26:52.726374054 ProcessGroupNCCL.cpp:1375] [PG 0 (default_pg) Rank 3] First PG on this rank that detected no heartbeat of its watchdog.
[rank0]:[E622 01:26:52.733272608 ProcessGroupNCCL.cpp:1413] [PG 0 (default_pg) Rank 0] Heartbeat monitor timed out! Process will be terminated after dumping debug info. workMetaList_.size()=1
[rank2]:[E622 01:26:52.733276826 ProcessGroupNCCL.cpp:1413] [PG 0 (default_pg) Rank 2] Heartbeat monitor timed out! Process will be terminated after dumping debug info. workMetaList_.size()=1
[rank3]:[E622 01:26:52.733286283 ProcessGroupNCCL.cpp:1413] [PG 0 (default_pg) Rank 3] Heartbeat monitor timed out! Process will be terminated after dumping debug info. workMetaList_.size()=1
[rank3]:[F622 01:36:52.743006702 ProcessGroupNCCL.cpp:1224] [PG 0 (default_pg) Rank 3] [PG 0 (default_pg) Rank 3] ProcessGroupNCCL's watchdog got stuck for 600 seconds without making progress in monitoring enqueued collectives. This typically indicates a NCCL/CUDA API hang blocking the watchdog, and could be triggered by another thread holding the GIL inside a CUDA api, or other deadlock-prone behaviors.If you suspect the watchdog is not actually stuck and a longer timeout would help, you can either increase the timeout (TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC) to a larger value or disable the heartbeat monitor (TORCH_NCCL_ENABLE_MONITORING=0).If either of aforementioned helps, feel free to file an issue to PyTorch about the short timeout or false positive abort; otherwise, please attempt to debug the hang. workMetaList_.size() = 1
[rank0]:[F622 01:36:52.743044824 ProcessGroupNCCL.cpp:1224] [PG 0 (default_pg) Rank 0] [PG 0 (default_pg) Rank 0] ProcessGroupNCCL's watchdog got stuck for 600 seconds without making progress in monitoring enqueued collectives. This typically indicates a NCCL/CUDA API hang blocking the watchdog, and could be triggered by another thread holding the GIL inside a CUDA api, or other deadlock-prone behaviors.If you suspect the watchdog is not actually stuck and a longer timeout would help, you can either increase the timeout (TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC) to a larger value or disable the heartbeat monitor (TORCH_NCCL_ENABLE_MONITORING=0).If either of aforementioned helps, feel free to file an issue to PyTorch about the short timeout or false positive abort; otherwise, please attempt to debug the hang. workMetaList_.size() = 1
[rank2]:[F622 01:36:52.743294147 ProcessGroupNCCL.cpp:1224] [PG 0 (default_pg) Rank 2] [PG 0 (default_pg) Rank 2] ProcessGroupNCCL's watchdog got stuck for 600 seconds without making progress in monitoring enqueued collectives. This typically indicates a NCCL/CUDA API hang blocking the watchdog, and could be triggered by another thread holding the GIL inside a CUDA api, or other deadlock-prone behaviors.If you suspect the watchdog is not actually stuck and a longer timeout would help, you can either increase the timeout (TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC) to a larger value or disable the heartbeat monitor (TORCH_NCCL_ENABLE_MONITORING=0).If either of aforementioned helps, feel free to file an issue to PyTorch about the short timeout or false positive abort; otherwise, please attempt to debug the hang. workMetaList_.size() = 1
W0622 01:37:23.259000 139910449522496 torch/distributed/elastic/multiprocessing/api.py:858] Sending process 1288873 closing signal SIGTERM
W0622 01:37:23.264000 139910449522496 torch/distributed/elastic/multiprocessing/api.py:858] Sending process 1288875 closing signal SIGTERM
W0622 01:37:53.265000 139910449522496 torch/distributed/elastic/multiprocessing/api.py:875] Unable to shutdown process 1288875 via 15, forcefully exiting via 9
E0622 01:37:53.711000 139910449522496 torch/distributed/elastic/multiprocessing/api.py:833] failed (exitcode: -6)

[yifuwang]

Hmm I was able to repro. You are right that speculative decoding + INTRA_NODE_COMM=1 can result in a hang after certain iterations. I poked around a bit and had the following observations:

• For me it always hung at a specific iteration
• The problem goes away if I call manual_seed at the beginning of each iteration

These make me think that the issue is data depedent. Though I don't have a lead on what it could be. Need to spend more time digging.