Remove no exp usage from logical rule Part I

[NuojCheng]

Description

This PR deprecates

• activation_batch_no_exp
• activation_length_no_exp

from logical names.

After this change

• activation_batch always includes "expert" physical axis
• activation_length does not include "expert"

activation_batch_moe and activation_batch_no_exp_moe are kept unchanged in moe.py.

Other logical names containing "_no_exp" will be deprecated in following PR.

Tests

CI tests.

Inference run v5p-8

NEW_MODEL_DESIGN=1 python src/maxtext/inference/vllm_decode.py src/maxtext/configs/post_train/rl.yml model_name=qwen3-30b-a3b tokenizer_path=Qwen/Qwen3-30B-A3B ici_tensor_parallelism=4 ici_expert_parallelism=1 enable_dp_attention=false hbm_utilization_vllm=0.3 load_parameters_path=gs://parambole-qwen3-moe-verification/unscanned/qwen3-30b-a3b-thinking-2507/14_08_2025/0/items vllm_hf_overrides='{architectures: ["MaxTextForCausalLM"]}' prompt="Suggest some famous landmarks in London." 2>&1 | tee  qwen3_moe_vllm_0.log

Output: https://paste.googleplex.com/5103945832857600

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

❌ Patch coverage is 91.30435% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/maxtext/layers/decoders.py	66.66%	1 Missing ⚠️
src/maxtext/layers/nnx_decoders.py	50.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[RissyRan]

Could you help check if this breaks inference workflow? Potentially this has been deprecated due to vLLM migration.

[NuojCheng]

Could you help check if this breaks inference workflow? Potentially this has been deprecated due to vLLM migration.

AFAIK inference.yml is deprecated. I will check inference path in https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/configs/inference/vllm.yml.

[gobbleturk]

note this should be activation_batch and activation_length - not activation_batch_moe and activation_length_moe

its a bit subtle and I'm not sure how to educate everyone, but the _moe rules should only be used deep inside moe after the tokens have moved to their expert shard. At this point in the model/code they have not been routed to their expert shard yet, so we should use the non _moe

this actual tensor, the logits, never need to get routed so never need to use _moe shardings....

[gobbleturk]

Doesn't have to be changed for this PR I guess

[NuojCheng]

ye I set _moe for all logical names inside RouteMoE, which is wrong. Now I understand why.

[richjames0]

lgtm