'airsnal_run_pipeline' command fails to finish.

[morialo3]

Running airsnal_run_pipeline give an error when it reaches "point predictions":

Fitting player model for FWD ...
Points prediction for player Kevin De Bruyne
gameweek: 6 vs NFO at home
Expected points: 0.00
gameweek: 7 vs WOL away
Expected points: 0.00
gameweek: 8 vs ARS away
Expected points: 0.00
Points prediction for player Harry Kane
gameweek: 6 vs ARS away
Expected points: 0.00
gameweek: 7 vs LIV at home
Expected points: 0.00
gameweek: 8 vs LUT away
Expected points: 0.00
2023-09-19 13:21:22.286525: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_I2023-09-19 13:21:22.286584: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.286579: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
NITIALIZED: initialization error
2023-09-19 13:21:22.286723: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.288209: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.288729: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_I2023-09-19 13:21:22.288752: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
NITIALIZED: initialization error
Points prediction for player Trent Alexander-Arnold
gameweek: 6 vs WHU at home
Expected points: 0.00
gameweek: 7 vs TOT away
Expected points: 0.00
gameweek: 8 vs BHA away
Expected points: 0.00
2023-09-19 13:21:22.292415: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.292877: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
Points prediction for player Ivan Toney
gameweek: 6 vs EVE at home
Expected points: 0.00
gameweek: 7 vs NFO away
Expected points: 0.00
gameweek: 8 vs MUN away
Expected points: 0.00
2023-09-19 13:21:22.296673: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.296852: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.299751: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.424797: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.475967: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.487848: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error
2023-09-19 13:21:22.489218: F external/xla/xla/stream_executor/cuda/cuda_driver.cc:149] Failed setting context: CUDA_ERROR_NOT_INITIALIZED: initialization error

then the process just continues to just to stop and throw a trace without completing he task:

Traceback (most recent call last):
File "/home/amro/miniconda3/envs/airsenalenv/bin/airsenal_run_pipeline", line 8, in
sys.exit(run_pipeline())
^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/click/core.py", line 1157, in call
return self.main(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/click/core.py", line 1078, in main
rv = self.invoke(ctx)
^^^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/click/core.py", line 1434, in invoke
return ctx.invoke(self.callback, **ctx.params)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/click/core.py", line 783, in invoke
return __callback(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/airsenal/scripts/airsenal_run_pipeline.py", line 174, in run_pipeline
opt_ok = run_optimize_squad(
^^^^^^^^^^^^^^^^^^^
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/airsenal/scripts/airsenal_run_pipeline.py", line 287, in run_optimize_squad
run_optimization(
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/airsenal/scripts/fill_transfersuggestion_table.py", line 595, in run_optimization
fill_suggestion_table(baseline_score, best_strategy, season, fpl_team_id)
File "/home/amro/miniconda3/envs/airsenalenv/lib/python3.11/site-packages/airsenal/framework/optimization_utils.py", line 252, in fill_suggestion_table
best_score = best_strat["total_score"]
~~~~~~~~~~^^^^^^^^^^^^^^^
TypeError: 'NoneType' object is not subscriptable
Total progress: 5%|███▏ | 1/22 [03:59<1:23:47, 239.41s/it]

I'll attach the full result of airsnal_run_pipeline to this gist"
https://gist.github.com/morialo3/cd7fde8e69027b69ac472427cf4986ee

[nbarlowATI]

Hi @morialo3 , interesting - I haven't seen this before. I am guessing that Jax (which is used by numpyro in the points prediction) is somehow configured to try and use the GPU. I don't know why this would be, or if it might be possible (or even beneficial?) to run on the GPU, but a potential way to force it to use CPU could be to do something like:

import jax
jax.config.update("jax_default_device", jax.devices("cpu")[0])

in airsenal/framework/bpl_interface.py before the line from bpl import ExtendedDixonColesMatchPredictor...
You'd then need to rerun pip install ..
Could you try this and let us know if it works?

[morialo3]

I patched airsenal/framework/bpl_interface.py , and it works with no issues. I usually run airsnal_{update_db,run_prediction,run_optimization} individually, so I never encountered this issue, apparently jax needs some adjustments to work with the gpu offload part, any suggestions in that part? thanks for the quick response.

[jack89roberts]

I usually run airsnal_{update_db,run_prediction,run_optimization} individually, so I never encountered this issue

Based on this my guess is this is something to do with multiprocessing and jax/GPUs not playing together nicely. I believe airsenal_run_prediction defaults to running single-threaded, but airsenal_run_pipeline defaults to running on four threads.

So airsenal_run_pipeline --num_thread 1 might work for you, but that would be worse than Nick's fix above because the majority of AIrsenal's runtime is in the optimisation, and running the optimisation single-threaded will be much slower (and there's nothing in the optimisation code that would make use of a GPU).

We might just want to include the workaround to force jax to always use CPU by default in AIrsenal.