Kimi-K2.6 unscanned checkpoint converter OOMs on hosts < ~2.5 TB RAM (buffers all 64 shards before write)

[sardaniksd]

Summary

convert_deepseek_family_unscanned_ckpt --model_size kimi-k2.6-text accumulates every dequantized bf16 tensor for all 64 HF safetensors shards in a single in-memory chkpt_vars dict before assembling and writing Orbax. For Kimi-K2.6 that's ~2.3 TB of CPU RAM at peak. There's no streaming/low-memory mode, and the documented runbook doesn't mention a host-size requirement.

Environment

• maxtext==0.2.2 installed via pip install maxtext[tpu] (PyPI), Python 3.12, CPU-only torch==2.12.0+cpu (Note: maxtext[tpu] pulls torch==2.12.0+cu130 whose torch._dynamo segfaults at import on a CPU-only host — forced reinstall with --index-url https://download.pytorch.org/whl/cpu to fix.)
• Host: GCP n2-standard-32 (32 vCPU, 128 GB RAM), 2 TB pd-balanced disk, Ubuntu 22.04.
• Zone: us-east5-a (chosen to be local to v5e TPU quota in this project; us-east5 has no m1/m2/m3-ultramem SKUs).
• HF source: moonshotai/Kimi-K2.6 (96 files, 64 *.safetensors shards, 555 GB on disk).

Repro

python -m maxtext.checkpoint_conversion.standalone_scripts.convert_deepseek_family_unscanned_ckpt \
    --model_size kimi-k2.6-text \
    --base_model_path /home/$USER/kimi-k2.6-hf \
    --maxtext_model_path gs://<bucket>/k2.6-unscanned

Memory growth observed (RSS via free -g)

After loading shard	RSS used
2	16 GB
3	53 GB
4	90 GB
5	124 GB
6	SIGKILL (rc=137)

~35–37 GB per shard, linear. Extrapolating across all 64 shards → ~2.3 TB RSS at peak.

Root cause (from reading the script)

In convert_deepseek_family_unscanned_ckpt.py, the outer loop over ckpt_paths populates a single chkpt_vars = {} with every dequantize_pack_quantized_int4(...) result before the assembly phase begins. Nothing is freed; nothing is written until the entire model is resident.

Asks

1. Is there a supported low-memory / streaming mode for the K2.6 (or any K2-Thinking/K2.5) converter? The K2.6 runbook doesn't mention host-memory requirements.
2. If not, what host SKU does the MaxText team use to convert K2.6? m1-ultramem-160 (3.75 TB) appears necessary but is unavailable in zones where v5e Lite TPU quota lives (us-east5, europe-west4).
3. Would a per-layer streaming patch be welcome upstream? Happy to PR it if there's interest.

Workarounds considered

• Cross-region: convert in us-central1 (has m1-ultramem-160), pay ~$80 GCS egress to land the Orbax in us-east5.
• Patch the script to dequant + emit per layer, freeing after each.
• Use the scanned converter then unscan at decode time — does this avoid the same accumulation?

[discobot]

The buffering you found is only half the peak: assembly holds a second full fp16 copy,
and the default save path re-materializes the whole pytree in RAM via
shard_jax_weights, so a loader-only patch would still OOM at save time. Fix in
#4160: shard loading is now lazy by default, and --low_memory true stages converted
weights in disk-backed memmaps and saves via the single-device path (bit-identical,
topology-independent checkpoint), keeping peak RSS at O(one tensor).