How to generate workloads in orojenesis

[wangxdgg]

Hi, I'm currently using orojenesis and working with the multi-layer test script orojenesis_multi.ipynb. However, it seems to require workloads input files which are missing. Could you please advise how to obtain the corresponding input files?

[hqjennynv]

[like] Jenny Huang reacted to your message:

…

________________________________ From: Angshuman Parashar ***@***.***> Sent: Thursday, May 29, 2025 9:29:34 PM To: NVlabs/timeloop ***@***.***> Cc: Jenny Huang ***@***.***>; Assign ***@***.***> Subject: Re: [NVlabs/timeloop] How to generate workloads in orojenesis (Issue #311) Assigned #311<#311> to @hqjennynv<https://github.com/hqjennynv>. — Reply to this email directly, view it on GitHub<#311 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AZ63TNGM2YQ2NGJFQ2YPXKD3A5335AVCNFSM6AAAAAB6FCBPRKVHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMJXHA3TKNBVHA2TIMQ>. You are receiving this because you were assigned.Message ID: ***@***.***>

[wangxdgg]

---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
Cell In[4], line 11
      7 force_rerun = False
     10 # Generate baseline optimal unconstrained unfused mappings.
---> 11 gen_mappings.gen_mappings(workload_dir, config_dir, output_dir, \
     12                             model_name=model_name, batch=batch_size, num_heads=num_heads, spatial_factor=None, arch_prefix=arch_prefix, ffmt=False, force_rerun=force_rerun)
     14 # Generate mappings following fusion friendly mapping templates (FFMT)
     15 gen_mappings.gen_mappings(workload_dir, config_dir, output_dir, \
     16                           model_name=model_name, batch=batch_size, num_heads=num_heads, spatial_factor=None, arch_prefix=arch_prefix, ffmt=True, force_rerun=force_rerun)

File /home/new_timeloop/timeloop/orojenesis/src/gen_mappings.py:205, in gen_mappings(workload_dir, config_dir, output_dir, model_name, batch, num_heads, ffmt, ar, nored, spatial_factor, arch_prefix, force_rerun)
    203     pair_chains_nc = utils.parse_yaml(graph_dir_nc / "pair_chains.yaml")
    204 else:
--> 205     pair_chains = utils.parse_yaml(graph_dir / "opt_schedules.yaml")
    206     pair_chains_nc = utils.parse_yaml(graph_dir_nc / "opt_schedules.yaml")
    207 layers_dict = utils.parse_yaml(graph_dir / "layers_dict.yaml")

File /home/new_timeloop/timeloop/orojenesis/src/utils.py:38, in parse_yaml(yaml_path)
     37 def parse_yaml(yaml_path):
---> 38     with open(yaml_path, "r") as f:
     39         data = yaml.full_load(f)
     40         return data

FileNotFoundError: [Errno 2] No such file or directory: 'workloads/gpt3-6.7b_graph_b16_matmul_nored/opt_schedules.yaml'

[hqjennynv]

Hi, sorry for missing the important workload files, I think we failed to push them because they are oversized for git.

I have pushed the code to generate the workload to the oaves_keep_max branch.
If you run the following commands, it should generate the corresponding workload files.

git checkout oaves_keep_max
cd orojenesis
python3 -m src.gen_gpt_graph  
python3 -m src.gen_gpt_graph  --matmul-only
python3 -m src.gen_graph_yaml
python3 -m src.gen_graph_yaml  --matmul-only

Please let us know if you have run into any other issues. Thanks!

[tuvaluhk]

Hi, I'm also trying to run orojenesis's example recently, and I generate workloads using:

cd orojenesis
python3 -m src.gen_gpt_graph  
python3 -m src.gen_gpt_graph  --matmul-only
python3 -m src.gen_graph_yaml
python3 -m src.gen_graph_yaml  --matmul-only

After that I try running the orojenesis_multi examples, however, I encounter another FileNotFoundError:

FileNotFoundError: [Errno 2] No such file or directory: 'workloads/attn-block_graph_b16_matmul_nored/opt_schedules.yaml'

Am I missing some extra command lines or the attn-block_graph_b16_matmul_nored is not uploaded?

[tuvaluhk]

OK, I kind of figure it out, there may be a miss on setting the model_name with template attn-block. One may set a specfied attention block description here.
I change model_name to the default gpt3-6.7b.