add main model card

README.md

@@ -0,0 +1,121 @@
+---
+license: other
+language:
+- en
+- zh
+tags:
+- qwen3.6
+- dsv4pro
+- glm
+- sft
+- rl
+- coding
+- gguf
+- fp8
+- mtp
+base_model: Merkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill
+---
+
+# Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding
+
+This is the Hugging Face main repository for the 27B dense coding model.
+
+- BF16 weights are stored at the repository root.
+- GGUF artifacts are stored under `GGUF/`, copied from the standalone GGUF repository.
+- FP8 artifacts are stored under `FP8/` when synced.
+- Standalone GGUF repository: [nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF](https://huggingface.co/nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF)
+
+The model is based on `Merkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill` and was further trained for executable coding, repair loops, final-answer discipline, and robust verification behavior.
+
+## Repository Layout
+
+| Path | Content |
+|---|---|
+| root | BF16 safetensors, config, tokenizer and chat template |
+| `GGUF/Q8_0/` | Q8 GGUF baseline, official MTP draft sidecar, eval summaries and manifest |
+| `GGUF/Q5_LynnStyle/` | 32GB recommended GGUF quantization, LynnStyle Dense protection rules and eval summaries |
+| `GGUF/Q4_LynnStyle/` | 24GB GGUF quantization, LynnStyle Dense protection rules and eval summaries |
+| `FP8/` | ModelOpt FP8 weights, SGLang script, MTP/NEXTN smoke and TPS data |
+
+GGUF files are multi-shard models. Load the first shard; llama.cpp will load the remaining shards from the same directory.
+
+## llama.cpp GGUF Example
+
+```bash
+llama-server \
+  -m GGUF/Q5_LynnStyle/Q5-imatrix-MTP-00001-of-00004.gguf \
+  -ngl 999 --split-mode layer \
+  -c 32768 -b 1024 -ub 256 -np 4 -cb \
+  --host 0.0.0.0 --port 8080 \
+  --jinja --reasoning on --reasoning-format deepseek
+```
+
+Optional speculative decoding with the MTP draft sidecar:
+
+```bash
+llama-server \
+  -m GGUF/Q5_LynnStyle/Q5-imatrix-MTP-00001-of-00004.gguf \
+  --model-draft GGUF/Q5_LynnStyle/Q5-imatrix-MTP-draft.gguf \
+  -ngl 999 --split-mode layer \
+  -c 32768 -b 1024 -ub 256 -np 4 -cb \
+  --host 0.0.0.0 --port 8080 \
+  --jinja --reasoning on --reasoning-format deepseek
+```
+
+For Q4 on 24GB GPUs, long-context use should prioritize the main model and KV cache. Only load the MTP sidecar for short-context, low-concurrency runs with enough spare VRAM.
+
+## FP8 / SGLang
+
+Recommended default: no MTP. MTP/NEXTN is supported but did not improve throughput in the current short-output smoke setup.
+
+```bash
+bash FP8/scripts/sglang_fp8_start.sh
+MTP=1 bash FP8/scripts/sglang_fp8_start.sh
+```
+
+Measured on RTX PRO 6000 Blackwell Server Edition with SGLang `0.5.14`, context `16384`, 8 short coding requests per mode:
+
+| Config | C1 | C2 | C4 |
+|---|---:|---:|---:|
+| no-MTP | 20.7581 tok/s | 32.6418 tok/s | 61.8252 tok/s |
+| MTP/NEXTN n=2, draft tokens=3 | 18.3805 tok/s | 26.1985 tok/s | 47.0971 tok/s |
+
+## Q8 Clean Reference Scores
+
+FP8 quality scores have not been independently rerun. For FP8, the table below is a Q8 clean-score reference for the same training branch.
+
+| Benchmark | Score |
+|---|---:|
+| MMLU500 | 462/500 = 92.4% |
+| GPQA198 | 162/198 = 81.82% |
+| LBC100 | 78/100 |
+| Coding100 current clean | 88/100 |
+
+## Quantized Clean Results
+
+| Version | MMLU500 | GPQA198 | LBC100 | Coding100 |
+|---|---:|---:|---:|---:|
+| Q8 | 92.4% | 81.82% | 78/100 | 88/100 |
+| Q5 LynnStyle | 92.2% | 78.79% | 75/100 | not rerun |
+| Q4 LynnStyle | 92.0% | 74.75% | 71/100 | 82/100 |
+
+Clean-gate policy: harness pollution, server-not-ready rows, parse failures caused by infrastructure, partial rows, and connection failures are not counted as model errors.
+
+## LynnStyle Dense Quantization
+
+This 27B model is dense, not MoE. LynnStyle Dense quantization protects critical layer positions, full-attention layers, attention gates, SSM tensor families, and selected tensor families while compressing ordinary regions more aggressively. Low-bit releases use explicit protection plus imatrix or equivalent importance calibration.
+
+---
+
+# 中文说明
+
+这是 27B Dense 编程模型的 Hugging Face 主仓。
+
+- 根目录放 BF16 权重。
+- `GGUF/` 放从 GGUF 独立仓复制过来的 Q8/Q5/Q4/Q3 等 GGUF 运行产物。
+- `FP8/` 放 FP8 权重、SGLang 启动脚本和 TPS/MTP smoke 数据。
+- GGUF 独立仓：[nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF](https://huggingface.co/nerkyor/Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF)
+
+本模型基于 `Merkyor/Qwen3.6-27B-DSV4Pro-Thinking-Distill` 继续后训练，重点强化可执行代码、闭环修复、验证纪律、final-answer 可靠性和推理预算控制。
+
+GGUF 多分片模型只需要指定第一片；MTP sidecar 是可选 speculative decoding 加速件，不是主模型。FP8 默认推荐 no-MTP；当前短输出压测下 no-MTP TPS 高于 MTP/NEXTN。