Skip to content

Conversation

@eltsai
Copy link
Collaborator

@eltsai eltsai commented Dec 16, 2025

We are running into an OOM error:

jax.errors.JaxRuntimeError: INTERNAL: RET_CHECK failure (platforms/xla/service/jellyfish/bounds_check.cc:708) allocation_size_words <= std::numeric_limits<int32_t>::max() 4180377600 %s1 = inlined_call_operand.hbm [shape: u32[64,75600,13824], index: 1, kind: output, shape index: {}] /* operand 1 / %5539 = dma.vmem_to_hbm [thread:$0] (%p7303), /vmem=/%s5537, /size_in_granules=/%s7302, /hbm=/%s5532, /dst_syncflagno=/%s5515 /.

if we set jax_default_prng_impl as unsafe_rbg and use --xla_tpu_spmd_rng_bit_generator_unsafe=true flag, it breaks the sharding of the mlp block in the transformer and caused OOM error.

This PR brings back the reverted optimizations and disable the unsafe rng flag.

The run with the flag disabled run successfully: log

@github-actions
Copy link

@eltsai eltsai requested a review from coolkp December 16, 2025 18:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant