Prefix caching | Mamba memory only.

[lmcafee-nvidia]

Summary

• Hybrid models (Transformer + Mamba) cannot skip prefill computation because Mamba layers maintain recurrent states that depend on the full sequence history, unlike attention KV cache blocks which are self-contained and reusable.
• This PR adds a guard in _compute_prefix_match that forces prefix_skip_tokens = 0 when is_hybrid_model is True, so matched prefix blocks are still shared (saving memory) but all prompt tokens are still processed through the model (preserving Mamba state correctness).
• Adds 4 tests in TestHybridModelMemoryOnly verifying: no prefill skipping, block reuse for memory savings, correct ref counts for shared blocks, and all prompt tokens present in context.

Details

When prefix caching is enabled for a hybrid model, the system operates in "memory-only" mode:

• KV cache blocks are shared across requests with matching prefix hashes, reducing memory consumption.
• Prefill is NOT skipped because Mamba layers must process the full sequence to reconstruct their internal states.

The change is a single 3-line guard in _compute_prefix_match (~line 1624 of dynamic_context.py):

# Hybrid models: disable prefill skipping (no Mamba states per block),
# but keep matched blocks for memory sharing.
if self.is_hybrid_model:
    prefix_skip_tokens = 0

Benchmarked on a 2B hybrid model (23 Mamba + 4 Attention + 23 MLP layers, 50 total) with 10 identical requests (644 tokens each):

• 64.2% block savings (31.0 → 11.1 blocks used)
• 0% prefill token reduction (6440 tokens in all configs), as expected
• Token-for-token output correctness vs. prefix caching disabled

Test plan

• test_no_prefill_skipping_for_hybrid_model: verifies prefix_skip_tokens == 0 and effective_chunk_length == chunk_length even when blocks match
• test_matched_blocks_reused_saving_memory: verifies second request consumes no additional blocks from pool
• test_ref_counts_incremented_for_matched_blocks: verifies matched blocks have ref_count == 2 after sharing
• test_all_prompt_tokens_in_context: verifies all prompt tokens are active (none skipped) and kv_length_offset == 0

/opt/venv/bin/python -m torch.distributed.run --nproc-per-node 1 -m pytest \
  tests/unit_tests/inference/contexts/test_dynamic_prefix_caching.py -v

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[lmcafee-nvidia]

/ok to test 13af54a

[lmcafee-nvidia]

/ok to test 13af54a

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/22643519147

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/22645678779