sprapp/eagle-grpo-iter19-q4k-uniform-noimatrix-enc

eagle-grpo-iter19 — I-Quality pack, uniform Q4_K, NO imatrix

READ THIS BEFORE COMPARING THIS PACK TO ANY iter_267 NUMBER.

What this is

An I-Quality .iqpt pack of the Oaica V4-Flash iter19 checkpoint (SFT + GRPO final, DeepSeek-V4-Flash 284B, 43 layers, 256 routed experts/layer), produced by pipeline/iquality/pack_cbalanced_proposed.py --preset c-balanced-proposed.

Files are AES-256-CTR encrypted (one random IV per file). The manifest mapping original_path -> encrypted_name -> IV is NOT in this repo — it is held separately. Without it these files are undecryptable.

What this is NOT: it is not validated c-balanced-proposed

The c-balanced-proposed recipe assigns routed experts across a hot/mid/cold spread (Q6_K/Q5_K hot, Q4_K mid, IQ3/IQ2_S cold) driven by a per-expert activation imatrix. That imatrix was computed on the bitdeer box and did not survive its destruction; no copy exists on USB, on a100b, or on HF.

With no --expert-scores, every routed expert receives the flat default score 0.5, and tier_policy.tier_for's band logic (s >= 1 - hot_frac - mid_frac → 0.5 >= 0.5) places all 11,008 experts in the mid band → `Q4_K`. Nothing reaches the IQ cold tiers.

So this pack is that recipe degraded to uniform Q4_K. Consequences:

Uniform Q4K was chosen deliberately over shrinking via the band fractions: `IQ3`/`IQ2S` are I-quants whose quality depends on imatrix-weighted codebook selection, so without an imatrix they degrade substantially more than K-quants do. Smaller and disproportionately worse was not a good trade. K-quants tolerate missing sensitivity data far better.

Do not benchmark this against the 117 GB iter_267 figures and conclude something regressed. The difference is the missing imatrix, not a change in the packer, the tiers, or the model.

If an iter19 activation imatrix is ever computed, repacking to the true mixed recipe is the obvious follow-up and should both shrink the pack and improve quality-per-byte.

Measured numbers (real, not estimated)

• —input: 268 GB HF-native fp8 safetensors (1479 shards, block-128 e4m3 + fp32 weight_scale_inv)
• —per layer: 13.16 GB in → 3.76 GB out, compression 3.50×
• —total: ~162 GB (43 layers + globals)
• —RMS vs source: expert 0.0023, attn 0.00061, shared 0.00009, router 0.0002
• —RMS-monotone guard: PASS

Serving

.iqpt is the packer's container; prism serves `.iqbin`. Convert with pipeline/iquality/iqpt_to_iqbin.py (torch-side pickle → little-endian binary; dequant math lives in Rust). .iqbin is the same size as .iqpt.

The prism-server binary must be built from the `prism-perf-deploy` branch — that is where V4FlashModel / v4flash.rs and the deepseek_v4 iqbin path live. Older prism builds (e.g. a100b /workspace/prism-server, Jul 23) have no iqbin support and cannot load this.

Provenance

• —source checkpoint: perahu_backup/eagle_grpo_hf_iter19_fp4/ (plaintext, USB)
• —packer: pipeline/iquality/pack_cbalanced_proposed.py, SafetensorsSource (O_DIRECT reader)
• —verified by: pipeline/iquality/verify_iqpt.py
• —license: Proprietary — All Rights Reserved.