Add default compiler options for v6e #887
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
samos123
force-pushed
the
v6e-compiler-options-clean
branch
from
564455c to
7ca21ff
Compare
Ethanlm
reviewed
Dec 11, 2024
Ethanlm
reviewed
Dec 11, 2024
samos123
force-pushed
the
v6e-compiler-options-clean
branch
from
7ca21ff to
655c539
Compare
Ethanlm
reviewed
Dec 12, 2024
ruomingp
approved these changes
Dec 12, 2024
Ethanlm
reviewed
Dec 12, 2024
| # Improved performance for v6e. | ||
| xla_tpu_scoped_vmem_limit_kib=98304, | ||
| xla_tpu_enable_async_collective_fusion=True, | ||
| xla_tpu_enable_async_collective_fusion_fuse_all_gather=True, |
There was a problem hiding this comment.
Needs These need to be "true" instead of True, otherwise it will be converted to 1 by xla_flags_from_options and then fail
I1212 21:41:11.645763 132689403348992 launch.py:112] LIBTPU_INIT_ARGS='--xla_tpu_spmd_rng_bit_generator_unsafe=1 --xla_tpu_enable_latency_hiding_scheduler=true --xla_tpu_perform_spmd_cse_prevention=false --megascale_grpc_premap_memory_bytes=17179869184 --xla_tpu_scoped_vmem_limit_kib=98304 --xla_tpu_enable_async_collective_fusion=1 --xla_tpu_enable_async_collective_fusion_fuse_all_gather=1 --xla_tpu_enable_async_collective_fusion_multiple_steps=1 --xla_tpu_overlap_compute_collective_tc=1 --xla_enable_async_all_gather=1 --xla_tpu_enable_all_experimental_scheduler_features=1 --xla_tpu_enable_scheduler_memory_pressure_tracking=1 --xla_tpu_host_transfer_overlap_limit=24 --xla_tpu_aggressive_opt_barrier_removal=1 --xla_lhs_prioritize_async_depth_over_stall=1 --xla_tpu_enable_ag_backward_pipelining=1 --xla_should_allow_loop_variant_parameter_in_chain=1 --xla_should_add_loop_invariant_op_in_chain=1 --xla_max_concurrent_host_send_recv=100 --xla_tpu_scheduler_percent_shared_memory_limit=90 --xla_latency_hiding_scheduler_rerun=2 --xla_tpu_use_enhanced_launch_barrier=1'
2024-12-12 21:41:14.669321: I external/tsl/tsl/platform/default/grpc_credentials.cc:30] gRPC insecure client credentials are used.
I1212 21:41:14.670917 132689403348992 distributed.py:119] Connecting to JAX distributed service on ethanli-fuji-70b-v2-test1-job-0-0.ethanli-fuji-70b-v2-test1:8476
2024-12-12 21:41:14.999583: I external/xla/xla/pjrt/distributed/client.cc:135] Connected to distributed JAX controller
ERROR: Illegal value '1' specified for flag 'xla_tpu_enable_async_collective_fusion_fuse_all_gather'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_enable_async_all_gather'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_tpu_enable_scheduler_memory_pressure_tracking'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_tpu_aggressive_opt_barrier_removal'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_lhs_prioritize_async_depth_over_stall'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_should_allow_loop_variant_parameter_in_chain'; expected one of true/enabled, false/disabled or auto
ERROR: Illegal value '1' specified for flag 'xla_should_add_loop_invariant_op_in_chain'; expected one of true/enabled, false/disabled or auto
There was a problem hiding this comment.
that's what I had originally, but pytype doesn't like that and will fail. So I see 2 options:
- Change behavior of axlearn that converts True boolean to 1 and let it return "true" instead.
- Ignore pytype check
kelvin-zou
reviewed
Dec 21, 2024
| xla_tpu_use_enhanced_launch_barrier=True, | ||
| # Sparsecore offloading for all reduce. | ||
| # Uncomment below flags to enable it. | ||
| # xla_sc_disable_megacore_partitioning=True, |
There was a problem hiding this comment.
@Ethanlm do you recall why you commented all them off?
There was a problem hiding this comment.
We weren't sure if this would perform better for everyone. It's a newer feature as well. So Ethan recommended not enabling by default. For both llama 2 70b and 405b, we saw much better performance with it being enabled though.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.