Initialize repo: README, gitignore, LFS tracking

.gitattributes

@@ -0,0 +1 @@
+*.safetensors filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,4 @@
+docker-compose.yml
+profiles/
+tunableop/
+openai-infqps-*.json

README.md

@@ -0,0 +1,128 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: image-text-to-text
+language:
+- en
+tags:
+- quantized
+- rfi
+- mixed-precision
+base_model:
+- migtissera/Tess-4-27B
+base_model_relation: quantized
+---
+
+> [!IMPORTANT]
+> ## RFI/RFA hybrid quant of [migtissera/Tess-4-27B](https://huggingface.co/migtissera/Tess-4-27B)
+>
+> **Runtime:** requires [`tcclaviger/vllm:latest`](https://hub.docker.com/r/tcclaviger/vllm) — an **RDNA 4 (gfx12xx)** vLLM image and the only build with the RFI/RFA kernels; no other vLLM build loads these weights. **Not validated on any other hardware at this time.**
+
+# Tess-27B-RFI
+
+Mixed-precision (RFI/RFA hybrid) quantization of **Tess-4-27B** by Migel Tissera —
+an agentic, thinking-native finetune of Qwen3.6-27B. All credit for the model to its
+author; this repo only changes the numerics.
+
+## Quantization by component
+
+- **Attention and all other linear layers** — 8-bit integer weights (RFI): group size 32, symmetric, Hadamard-32 rotation, block-float scales stored as int8 mantissa + int8 exponent, int8 activation compute path.
+- **MLP layers** — 4-bit float weights (RFA): IQ4_NL non-linear grid, group size 16, asymmetric, Hadamard-16 rotation, the same block-float int8 scale encoding.
+- **Kept in bf16 (not quantized)** — vision encoder, linear-attention (GDN) blocks, MTP speculative-decode layers, embeddings, and the lm_head.
+
+## Evaluation results
+
+Four builds measured with identical methodology, each against its own live vLLM endpoint (July 2026):
+**Qwen3.6-27B** (bf16 base) → **Tess-4-27B** (the tune, bf16) → **Tess-27B-RFI** (this quant),
+with **Qwen3.6-35B-A3B** (MoE, bf16) as a comparative reference point.
+
+### Tune impact — Qwen3.6-27B → Tess-4-27B
+
+| Metric | Change |
+|---|---|
+| WikiText-2 perplexity | 7.060 → 6.669 (**−5.5%**) |
+| Codeneedle overall recall | 92.0% → 97.7% (**+5.8 pp**) |
+| MC accuracy (4 tasks) | ≈ flat (−0.1 to +1.8 pp) |
+| Tool-eval (full 69, TC-61 excl) | 89 → 85 (−4) |
+| GSM8K / MMLU (50q each) | 98→94% / 74→76% |
+| Decode @ conc 1 (ISL 128) | 63.7 → 59.9 tok/s (−6%) |
+| Decode @ conc 50 (ISL 128) | 541 → 528 tok/s (−2%) |
+
+### Quantization cost (RFI/RFA hybrid) — Tess-4-27B → Tess-27B-RFI
+
+| Metric | Change |
+|---|---|
+| WikiText-2 perplexity | 6.669 → 6.663 (≈ flat) |
+| Codeneedle overall recall | 97.7% → 98.0% (≈ flat) |
+| MC accuracy (4 tasks) | ≈ flat (−0.4 to +0.1 pp) |
+| Tool-eval (full 69, TC-61 excl) | 85 → 87 (**+2**) |
+| GSM8K / MMLU (50q each) | 94→98% / 76→82% (**both up**) |
+| Decode @ conc 1 (ISL 128) | 59.9 → 75.3 tok/s (**+26%**) |
+| Decode @ conc 50 (ISL 128) | 528 → 453 tok/s (−14%) |
+
+### Quality
+
+| Metric | Qwen3.6-27B (base) | Tess-4-27B | **Tess-27B-RFI** | Qwen3.6-35B-A3B |
+|---|---|---|---|---|
+| WikiText-2 PPL (n_ctx 2048, lower is better) | 7.0597 | 6.6691 | **6.6632** | 6.5092 |
+| ARC-Challenge (acc_norm) | 59.13% | 60.84% | **60.41%** | 55.20% |
+| ARC-Easy (acc_norm) | 75.76% | 77.53% | **77.40%** | 71.13% |
+| Winogrande (acc) | 76.87% | 77.43% | **77.51%** | 73.40% |
+| HellaSwag (acc_norm) | 84.04% | 84.21% | **84.27%** | 82.95% |
+
+Multiple-choice accuracy is lm-eval loglikelihood scoring, 0-shot.
+
+### Long-context positional recall (codeneedle)
+
+Verbatim function recall under 10K–80K-token contexts.
+
+| Corpus | Qwen3.6-27B (base) | Tess-4-27B | **Tess-27B-RFI** | Qwen3.6-35B-A3B |
+|---|---|---|---|---|
+| Python | 92.73% | 100% | **100%** | 99.09% |
+| C++ | 92.81% | 98.12% | **98.44%** | 98.44% |
+| Rust | 95.00% | 99.69% | **99.69%** | 99.38% |
+| JS (~80K tokens) | 87.50% | 93.13% | **93.75%** | 92.19% |
+| Tools | 96.52% | 99.57% | **99.57%** | 93.48% |
+| **Overall recall** | 91.95% | 97.73% | **97.97%** | 97.28% |
+
+### Tool calling & accuracy benches
+
+| Bench | Qwen3.6-27B (base) | Tess-4-27B | **Tess-27B-RFI** | Qwen3.6-35B-A3B |
+|---|---|---|---|---|
+| tool-eval final (full 69, TC-61 excl) | 89 | 85 | **87** | 90 |
+| GSM8K (50q) | 98.0% | 94.0% | **98.0%** | 96.0% |
+| MMLU (50q) | 74.0% | 76.0% | **82.0%** | 64.0% |
+| IFEval (20 prompts, prompt-level) | 90.0% | 90.0% | **90.0%** | 90.0% |
+
+### Decode throughput — tok/s output (ISL 128 / ISL 512)
+
+`vllm bench serve`, random dataset, OSL 128, saturation, 4× R9700 (gfx1201), TP 4.
+
+| Concurrency | Qwen3.6-27B (base) | Tess-4-27B | **Tess-27B-RFI** | Qwen3.6-35B-A3B |
+|---|---|---|---|---|
+| 1 | 63.7 / 65.2 | 59.9 / 63.8 | **75.3 / 69.3** | 91.9 / 114.4 |
+| 10 | 251.1 / 246.6 | 289.6 / 227.8 | **281.5 / 219.2** | 434.9 / 440.3 |
+| 25 | 390.3 / 336.8 | 492.0 / 341.1 | **429.1 / 284.0** | 688.9 / 563.7 |
+| 50 | 541.2 / 363.0 | 527.9 / 345.6 | **452.8 / 295.0** | 889.8 / 702.9 |
+
+## Notes
+
+All builds serve on the [tcclaviger/vllm:latest](https://hub.docker.com/r/tcclaviger/vllm) image, which has kernel tunes baked in.
+TunableOp is untuned — GEMMs run on default heuristic-determined values.
+
+## Credits
+
+- **Tess-4-27B** by Migel Tissera ([migtissera/Tess-4-27B](https://huggingface.co/migtissera/Tess-4-27B)) — the model quantized here:
+
+  ```bibtex
+  @misc{tissera2026tess4,
+    title        = {Tess-4-27B},
+    author       = {Migel Tissera},
+    year         = {2026},
+    howpublished = {\url{https://huggingface.co/migtissera/Tess-4-27B}}
+  }
+  ```
+
+- codeneedle (positional recall) originally by Alexander Ziskind, expanded test suite by tcclaviger (Rob Smith).
+- Tool-calling scenarios (incl. TC-61) run on tool-eval-bench by SeraphimSerapis (Tim Messerschmidt), scenario methodology adapted from ToolCall-15 by stevibe.
+- GSM8K, MMLU, and IFEval run via tool-eval-bench's built-in accuracy benchmarks at their defaults: GSM8K 8-shot CoT, MMLU 5-shot, IFEval zero-shot.