deepseek-v3-compact-proxy

DeepSeek-V3 Compact Proxy — Checkpoints

Trained checkpoints for deepseek-v3-compact-proxy, a mechanism-faithful proxy for DeepSeek-V3's architecture (Multi-head Latent Attention, DeepSeekMoE, Multi-Token Prediction) — not a reproduction of the 671B model. Every checkpoint here corresponds to a specific, documented experiment ("gate") in that repository; each gate's specification, implementation, tests, measured results, and interpretation live in its experiments/GATE_*.md file.

Checkpoints are hosted here rather than in the code repository because GitHub blocks any file over 100MB, and several of these (Configuration B, with optimizer state) are close to 2GB each.

Every checkpoint below is also reproducible from the exact command recorded in its gate document. This Hugging Face copy is a convenience rather than the only record.

Which one to use

compact_v3_wikitext103_2ep.pt is the best model here: 41.35 validation perplexity on WikiText-103, from 229,376,000 tokens (2 epochs, 64,000 steps). Download it together with tokenizer_wikitext103.json, which it was trained with. Decoding with a different tokenizer produces noise.

uv run huggingface-cli download syedazeez/deepseek-v3-compact-proxy `
  compact_v3_wikitext103_2ep.pt tokenizer_wikitext103.json --local-dir checkpoints

uv run python serve.py --checkpoint checkpoints\compact_v3_wikitext103_2ep.pt `
  --tokenizer checkpoints\tokenizer_wikitext103.json

That opens a browser playground at http://127.0.0.1:8000. This is a base model with no instruction tuning, so it continues text rather than answering questions.

The 15 Configuration A and B checkpoints below were trained on WikiText-2 with a different tokenizer, which is not uploaded here. Regenerate it by running v3_cli.py --real-corpus once, which fits it from the train split.

Configuration B, WikiText-103 (155M parameters, Gates V-W)

Both use tokenizer_wikitext103.json. Gate V's checkpoint is the shakedown run that Gate W continued from; it is kept because serve.py still names it as its default --checkpoint. Its 87.57 was re-measured over all 470 validation windows, against the 87.85 recorded in the gate document: the difference is the evaluation batch size, not the weights, because evaluate_tokens averages per-batch means and so weights the final partial batch like a full one.

Trained at batch 7, sequence 512, context 520, with MTP enabled and the loss weight annealed 0.3 to 0.1. The earlier 292.54 figure quoted for Gate Q is not comparable: it used WikiText-2 validation with a WikiText-2 tokenizer at context 264.

Configuration A (~15M parameters, Gates I-N)

Configuration B (~155M parameters, Gates O-Q)

Gate R (MLA weight absorption) is an inference-path code change validated by unit tests and benchmarks; it did not produce a training checkpoint.

Regenerating a checkpoint

Each gate document's "Experiment" section has the exact command. For example, Gate Q's best checkpoint:

uv run python v3_cli.py --real-corpus --steps 1954 --checkpoint-every 250 --eval-batches 32 --generate 32 --device cuda --enable-mtp --checkpoint checkpoints/compact_v3_wikitext_moe_scaleup_1m_gateq_mtp_anneal.pt