DeepSpeed
DeepSpeed copied to clipboard
microsoft
[BUG] File not found in autotuner cache in multi-node setting on SLURM
Describe the bug I am training an LLM using DeepSpeed and 12 nodes a 8 V100s per node. My training is generally working well (thanks DeepSpeed), but when I run multiple training runs in parallel, I run into trouble. I am getting these kinds of errors
Traceback (most recent call last):
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 473, in matmul_ext_update_autotune_table
fp16_matmul._update_autotune_table()
fp16_matmul._update_autotune_table()
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 450, in _update_autotune_table
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 450, in _update_autotune_table
TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 179, in _update_autotune_table
TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 179, in _update_autotune_table
cache_manager.put(autotune_table)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 98, in put
cache_manager.put(autotune_table)
os.rename(self.file_path + ".tmp", self.file_path)
FileNotFoundError: [Errno 2] No such file or directory: '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle.tmp' -> '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle'
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 98, in put
os.rename(self.file_path + ".tmp", self.file_path)
FileNotFoundError: [Errno 2] No such file or directory: '/gpfs/u/home/ANFM/ANFMbchl/scratch/.cache/Fp16Matmul_2d_kernel.pickle.tmp' -> '/gpfs/u/home/ANFM/ANFMbchl/scratch/.cache/Fp16Matmul_2d_kernel.pickle'
I thought that this is because maybe the directories are shared between the multiple runs, which can create race conditions.
My TMPDIR, TRITON_CACHE_DIR, and TORCH_EXTENSIONS_DIR are set as follows
export TMPDIR=$HOME/scratch/.cache
export TRITON_CACHE_DIR=$HOME/scratch/.cache
export TORCH_EXTENSIONS_DIR=$HOME/scratch/.cache/torch-extensions
To fix this, I tried to allocate one cache folder per run like so
export TMPDIR=$HOME/scratch/.cache
export TRITON_CACHE_DIR=$HOME/scratch/$SLURM_JOBID/.cache
export TORCH_EXTENSIONS_DIR=$HOME/scratch/$SLURM_JOBID/.cache/torch-extensions
mkdir -p $TRITON_CACHE_DIR
mkdir -p $TORCH_EXTENSIONS_DIR
but that also didn't work. Now I am getting this error
Traceback (most recent call last):
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 473, in matmul_ext_update_autotune_table
fp16_matmul._update_autotune_table()
fp16_matmul._update_autotune_table()
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 450, in _update_autotune_table
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 450, in _update_autotune_table
TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 179, in _update_autotune_table
TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 179, in _update_autotune_table
cache_manager.put(autotune_table)
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 98, in put
cache_manager.put(autotune_table)
os.rename(self.file_path + ".tmp", self.file_path)
FileNotFoundError: [Errno 2] No such file or directory: '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle.tmp' -> '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle'
File "/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 98, in put
os.rename(self.file_path + ".tmp", self.file_path)
FileNotFoundError: [Errno 2] No such file or directory: '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle.tmp' -> '/gpfs/u/home/ANFM/ANFMbchl/scratch/1167439/.cache/Fp16Matmul_2d_kernel.pickle'
ds_report output
[2024-06-12 03:08:15,154] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2024-06-12 03:08:15,765] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-devel package with yum
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
[WARNING] Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
No ROCm runtime is found, using ROCM_HOME='/opt/rocm-4.3.0'
[WARNING] NVIDIA Inference is only supported on Ampere and newer architectures
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.3
[WARNING] using untested triton version (2.3.0), only 1.0.0 is known to be compatible
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
runtime if needed. Op compatibility means that your system
meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-devel package with yum
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
[WARNING] Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
[WARNING] NVIDIA Inference is only supported on Ampere and newer architectures
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.3
[WARNING] using untested triton version (2.3.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch']
torch version .................... 2.3.0+cu121
deepspeed install path ........... ['/gpfs/u/scratch/ANFM/ANFMbchl/DeepSpeed/deepspeed']
deepspeed info ................... 0.14.3+488a823, 488a823, master
torch cuda version ............... 12.1
torch hip version ................ None
nvcc version ..................... 12.1
deepspeed wheel compiled w. ...... torch 2.4, cuda 12.1
shared memory (/dev/shm) size .... 377.69 GB
@jubueche - the related PR is now resolved, can you see if you are still hitting this if you use the latest DeepSpeed?
I am also experiencing this issue on a SLURM managed system
[2024-11-04 16:34:44,280] [INFO] [logging.py:96:log_dist] [Rank 0] Wrote metrics to /gpfs/home3/cbarkhof/transformers-deepspeed/autotuning_metric.json, /gpfs/home3/cbarkhof/transformers-deepspeed/autotuning_metric.json
0%| | 1/10000 [00:08<23:10:23, 8.34s/it]
0%| | 2/10000 [00:09<12:00:48, 4.33s/it]
0%| | 3/10000 [00:11<8:26:40, 3.04s/it]
0%| | 4/10000 [00:12<6:47:47, 2.45s/it]
0%| | 5/10000 [00:14<5:51:39, 2.11s/it]
0%| | 5/10000 [00:15<8:50:52, 3.19s/it]
Autotuning: done with running current ds config.
Exception ignored in atexit callback: <function matmul_ext_update_autotune_table at 0x14eba8baf920>
Traceback (most recent call last):
File "/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 477, in matmul_ext_update_autotune_table
fp16_matmul._update_autotune_table()
File "/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 454, in _update_autotune_table
TritonMatmul._update_autotune_table(__class__.__name__ + "_2d_kernel", __class__._2d_kernel)
File "/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 183, in _update_autotune_table
cache_manager.put(autotune_table)
File "/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/deepspeed/ops/transformer/inference/triton/matmul_ext.py", line 102, in put
os.rename(self.file_path + ".tmp", self.file_path)
FileNotFoundError: [Errno 2] No such file or directory: '/home/cbarkhof/autotune/.cache/Fp16Matmul_2d_kernel.pickle.tmp' -> '/home/cbarkhof/autotune/.cache/Fp16Matmul_2d_kernel.pickle'
[2024-11-04 16:33:47,323] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
runtime if needed. Op compatibility means that your system
meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-devel package with yum
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
[WARNING] Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
[WARNING] FP Quantizer is using an untested triton version (2.2.0), only 2.3.(0, 1) and 3.0.0 are known to be compatible with these kernels
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
[WARNING] gds requires the dev libaio .so object and headers but these were not found.
[WARNING] gds: please install the libaio-devel package with yum
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
gds .................... [NO] ....... [NO]
transformer_inference .. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.2
[WARNING] using untested triton version (2.2.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/torch']
torch version .................... 2.2.2+cu121
deepspeed install path ........... ['/gpfs/home3/cbarkhof/transformers-deepspeed/.venv/lib/python3.11/site-packages/deepspeed']
deepspeed info ................... 0.15.2, unknown, unknown
torch cuda version ............... 12.1
torch hip version ................ None
nvcc version ..................... 12.1
deepspeed wheel compiled w. ...... torch 0.0, cuda 0.0
shared memory (/dev/shm) size .... 377.81 GB
Is there any workaround for this?
@loadams I haven't tried yet and have just used my "workaround" since. @ClaartjeBarkhofTNO this works for me:
def put(self, table):
if self.file_path:
assert self.lock_path is not None
# with FileLock(self.lock_path):
# with open(self.file_path + ".tmp", 'wb') as handle:
# pickle.dump(table, handle)
# os.rename(self.file_path + ".tmp", self.file_path)
@loadams this is probably really bad, but it is surprisingly working. Would you know why?
@ClaartjeBarkhofTNO and @jubueche - thanks for the update on this, we will take a look.
Can you share your DeepSpeed version as well as any other info about your system you can share beyone using SLURM?