Ternarization_instructions_27b_en.md
| 1 | # Instructions for Ternarization and Export to GGUF/TQ2_0 — Qwen3.8-27B |
| 2 | |
| 3 | Model: `/mnt/nfs_share/Qwen3_8/`, class `JiRackDeltaNet_27b.py` |
| 4 | (3:1 hybrid Gated-DeltaNet / Gated-Attention, `model_type: qwen3_5`). |
| 5 | |
| 6 | ## Status (updated 2026-08-28) |
| 7 | |
| 8 | ✅ **Done:** |
| 9 | - HF → custom format conversion, full-pipeline testing |
| 10 | - **MTP block implemented and integrated** (see section 0 below) |
| 11 | - GGUF conversion successful (pt → safetensors → GGUF, bf16) |
| 12 | - Smoke test passed: model responds correctly ("Paris" for "What is the capital of France?") |
| 13 | - **Baseline perplexity: PPL = 5.4957 ± 0.12** (wikitext-100k, c=2048, bf16, lambda=0.0) |
| 14 | - File: `/mnt/nfs_share/Qwen3_8/wiki_100k.txt` (use this same file for comparison after ternarization) |
| 15 | - Log: `/mnt/nfs_share/Qwen3_8/ppl_baseline_bf16.log` (final line, 11 chunks) |
| 16 | |
| 17 | ⏳ **Next steps:** QAT warmup lambda 0.0→1.0, ternarization materialization, TQ2_0 export, final perplexity |
| 18 | |
| 19 | Current checkpoint: `/mnt/nfs_share/Qwen3_8/qwen38_27b_checkpoint_migrated.pt` (53.79GB, lambda=0.0) |
| 20 | - Contains MTP weights (transferred from the original HF checkpoint), but not yet ternarized. - It will be converted into a new `.pt` file with native MTP support using the `convert_qwen38_to_jirack.py` script (updated version). |
| 21 | |
| 22 | Quantization mechanism (same as the 32B model): `BitLinear` calculates gamma as the **per-tensor absmean**; |
| 23 | weights map to exactly three values (`-gamma, 0, +gamma`), ensuring a lossless TQ2_0 round-trip. |
| 24 | **Difference from the 32B model:** two architectural adjustments—see sections 0 and 1. |
| 25 | |
| 26 | --- |
| 27 | |
| 28 | ## 0. MTP Block (Medusa Token Prediction, layer 64) |
| 29 | |
| 30 | **What it is:** The new llama.cpp architecture (version ≥ ca3d5a3e1, tag b10665) requires |
| 31 | an MTP layer for models like `qwen3_5`. MTP is a speculative decoder (draft layer) |
| 32 | designed to improve batch generation, similar to Medusa. Implemented in `JiRackDeltaNet_27b.py`. **What has been done so far:** |
| 33 | - The `MTPBlock` class has been implemented in the code (15 tensors: fc, norms). |
| 34 | - MTP weights have been ported from the original Qwen3.8-27B HF checkpoint. |
| 35 | - They have been integrated into the GGUF file during conversion (`qwen38_27b_jirack_baseline_bf16.gguf`). |
| 36 | - MTP is **not** included in the ternarization process (it remains in full precision, ~100MB of weights). |
| 37 | |
| 38 | **Why we aren't ternarizing MTP:** |
| 39 | - The MTP layer is not essential for standard generation (it is used only when the speculative decoding flag is enabled). |
| 40 | - Ternarizing the 400 main layers is sufficient to achieve the target compression. |
| 41 | - Keeping MTP in full precision preserves the quality of the speculative decoder for users who choose to enable it. |
| 42 | |
| 43 | **For QAT warmup and materialization:** simply skip the `mtp.*` layers when applying |
| 44 | the `BitLinear` patch—this is already handled by `apply_bitlinear_patch(skip_mtp=True)`. |
| 45 | |
| 46 | --- |
| 47 | |
| 48 | ## 1. What gets ternarized and what doesn't (THREE EXCEPTIONS) |
| 49 | |
| 50 | By default (`include_gate_proj=False` in `apply_bitlinear_patch`), |
| 51 | the following 400 linear layers are patched: |
| 52 | - `linear_attn.in_proj_qkv`, `linear_attn.in_proj_z`, `linear_attn.out_proj` |
| 53 | - `self_attn.q_proj`, `k_proj`, `v_proj`, `o_proj` |
| 54 | - `mlp.gate_proj`, `up_proj`, `down_proj` |
| 55 | |
| 56 | **The following are NOT patched and must remain in full precision:** |
| 57 | - `linear_attn.in_proj_b` and `linear_attn.in_proj_a` — these are the gate scalars |
| 58 | for DeltaNet (dt/A-gate projections); they control recurrence stability. This was a deliberate decision: the risk of destabilizing the recurrence formula itself |
| 59 | through coarse ternary quantization was deemed greater than the potential |
| 60 | savings. |
| 61 | - All normalization layers (`RMSNorm`, `RMSNormGated`), `conv1d` (depthwise, |
| 62 | 4-tap), `dt_bias`, `A_log`. |
| 63 | - `embed_tokens`, `lm_head`. |
| 64 | |
| 65 | **Implication for GGUF export:** when quantizing to GGUF, you must explicitly |
| 66 | specify the tensor type for these layers (using `--tensor-type` or |
| 67 | the equivalent in your version of `llama-quantize`) so that |
| 68 | `in_proj_b`/`in_proj_a`, conv1d, and all normalization layers remain in f16/f32 |
| 69 | rather than being subjected to the general TQ2_0 setting during a full-file |
| 70 | quantization pass. Without explicit specification, some quantization scripts |
| 71 | apply a single type to all eligible tensors indiscriminately—which would |
| 72 | silently break the recurrence mechanism. |
| 73 | |
| 74 | ## 2. Architectural risk — ALREADY VERIFIED ✅ |
| 75 | |
| 76 | Gated DeltaNet is a hybrid architecture that is new to llama.cpp. The risk lay in the |
| 77 | fact that its support is less mature than that of Qwen2-attention in 32B models. **Verified (2026-08-28):** |
| 78 | - GGUF conversion (`pt` → `safetensors` → `qwen38_27b_jirack_baseline_bf16.gguf`): |
| 79 | successful, 866 tensors, 54.6GB |
| 80 | - Smoke test (llama-cli with chat-template): model responds correctly |
| 81 | ("The capital of France is Paris" with reasoning layer) |
| 82 | - Baseline perplexity on wikitext-100k: **PPL = 5.4957 ± 0.12** |
| 83 | (11 chunks of 2048 tokens, bf16, lambda=0.0) |
| 84 | |
| 85 | **Conclusion:** the architecture works correctly in llama.cpp. Any further |
| 86 | degradation during ternary/TQ2_0 quantization will stem from the quantization |
| 87 | process itself, not from implementation bugs. |
| 88 | |
| 89 | **Note for the future:** if perplexity spikes abnormally (>>6.5) during TQ2_0, |
| 90 | this is not an architectural risk—it indicates an issue with QAT warmup |
| 91 | or incorrectly excluded layers. |
| 92 | |
| 93 | ## 3. QAT training (when ternary quantization is required) |
| 94 | |
| 95 | 1. Load checkpoint: |
| 96 | ```python |
| 97 | model = JiRackQwen38ForCausalLM.load_checkpoint( |
| 98 | "/data/qwen38_27b_checkpoint_migrated.pt", dtype=torch.bfloat16) |
| 99 | ``` |
| 100 | 2. Gradually ramp up the warmup lambda from 0.0 to 1.0 (similar to the 32B model, |
| 101 | using `set_lambda` on all modules that support this method). |
| 102 | 3. Save the trained checkpoint. |
| 103 | |