We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
I tried to reproduce the experiment on a nvidia-a100-80g, but got the OOM issue, even for the 6.7b opt model.
python -m flexgen.flex_opt --model facebook/opt-6.7b --path DUMMY --percent 100 0 100 0 100 0 --overlap False --gpu-batch-size 64 --prompt-len 2048 --gen-len 2048 --hh-long-seq --hh-ratio 0.2 --hh-all 2048 model size: 12.386 GB, cache size: 128.000 GB, hidden size (prefill): 2.000 GB init weight... == after init weight == used: 126 GB, free: 1271 GB, cached: 370 GB, available: 1619 GB warmup - generate benchmark - generate == after init cache == used: 126 GB, free: 1271 GB, cached: 370 GB, available: 1619 GB Traceback (most recent call last): File "/opt/conda/envs/ptca/lib/python3.8/runpy.py", line 194, in _run_module_as_main return _run_code(code, main_globals, None, File "/opt/conda/envs/ptca/lib/python3.8/runpy.py", line 87, in _run_code exec(code, run_globals) File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 1423, in run_flexgen(args) File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 1318, in run_flexgen output_ids = model.generate( File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 942, in generate self.generation_loop_normal() File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 983, in generation_loop_normal self.compute_layer(i, j, k) File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 841, in compute_layer self.layers[j].forward(self.hidden[i][j][k], self.cache_read_buf[j][k], File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 482, in forward h, new_k_cache, new_v_cache, acc = self.compute.mha(h, mask, w_q, b_q, File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/pytorch_backend.py", line 365, in mha attn_weights = F.softmax(attn_weights, dim=2, dtype=torch.float32).to(torch.float16) File "/opt/conda/envs/ptca/lib/python3.8/site-packages/torch/nn/functional.py", line 1843, in softmax ret = input.softmax(dim, dtype=dtype) torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB (GPU 0; 79.17 GiB total capacity; 59.77 GiB already allocated; 17.68 GiB free; 60.53 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
The text was updated successfully, but these errors were encountered:
No branches or pull requests
I tried to reproduce the experiment on a nvidia-a100-80g, but got the OOM issue, even for the 6.7b opt model.
python -m flexgen.flex_opt --model facebook/opt-6.7b --path DUMMY --percent 100 0 100 0 100 0 --overlap False --gpu-batch-size 64 --prompt-len 2048 --gen-len 2048 --hh-long-seq --hh-ratio 0.2 --hh-all
2048
model size: 12.386 GB, cache size: 128.000 GB, hidden size (prefill): 2.000 GB
init weight...
== after init weight ==
used: 126 GB, free: 1271 GB, cached: 370 GB, available: 1619 GB
warmup - generate
benchmark - generate
== after init cache ==
used: 126 GB, free: 1271 GB, cached: 370 GB, available: 1619 GB
Traceback (most recent call last):
File "/opt/conda/envs/ptca/lib/python3.8/runpy.py", line 194, in _run_module_as_main
return _run_code(code, main_globals, None,
File "/opt/conda/envs/ptca/lib/python3.8/runpy.py", line 87, in _run_code
exec(code, run_globals)
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 1423, in
run_flexgen(args)
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 1318, in run_flexgen
output_ids = model.generate(
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 942, in generate
self.generation_loop_normal()
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 983, in generation_loop_normal
self.compute_layer(i, j, k)
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 841, in compute_layer
self.layers[j].forward(self.hidden[i][j][k], self.cache_read_buf[j][k],
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/flex_opt.py", line 482, in forward
h, new_k_cache, new_v_cache, acc = self.compute.mha(h, mask, w_q, b_q,
File "/scratch/AzureBlobStorage_INPUT1/vc_data_blob/users/yunanzhang/H2O/h2o_flexgen/flexgen/pytorch_backend.py", line 365, in mha
attn_weights = F.softmax(attn_weights, dim=2, dtype=torch.float32).to(torch.float16)
File "/opt/conda/envs/ptca/lib/python3.8/site-packages/torch/nn/functional.py", line 1843, in softmax
ret = input.softmax(dim, dtype=dtype)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 GiB (GPU 0; 79.17 GiB total capacity; 59.77 GiB already allocated; 17.68 GiB free; 60.53 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
The text was updated successfully, but these errors were encountered: