[tawei-amd]

Summary

qwen3.8-dflash2:27b: Qwen3.8-27B decoded with z-lab's DFlash 2 draft instead of the built-in MTP head.

• Catalog: tag qwen3.8-dflash2 / 27b, base_model: qwen3.8-mtp:27b, own package Qwen3.8-27B-DFlash2-NPU2 (weights only: dflash_weight.q4nx 1.39 GB, config.json, LICENSE), xclbin_folder: Qwen3.8-27B-NPU2 so it runs the base model's kernels (no xclbin copy).
• model_list.hpp: resolve_source always forced xclbin_folder to the dependant's own name; it now keeps one the entry sets. New catalog tests (explicit xclbin_folder; the shipped qwen3.8-dflash2 entry).
• Wrapper: Qwen3_8DFlash2 : Qwen3_8MTP (only names the draft), family qwen3.8-dflash2 registered on the same engine library (RegisterFlm<qwen3_8mtp_npu>), CI family tables.
• Engine library (libqwen3_8mtp_npu.so / .dll, XRT and HRX): from IRON PR AMD-ROCm-Internal/FastFlowLM_IRON#136, CI run 41346212324; sha256 below. The engine loads the draft package from draft_model_path and the DFlash2 drafter replaces the MTP head (FLM_DFLASH=0 keeps the MTP head).
• Docs: model card in qwen.md, decode benchmark in qwen3.8_results.md, notes in the request-parameters / prefix-cache / OpenAI pages. No new page, so no sidebar entry.

Verification

• Libraries are the IRON CI build, verified locally before vendoring: same outputs and speeds as the local build (10 prompts x 160 tokens, 16-case behaviour matrix).
• Decode (greedy, ten prompts): 3.74 tok/s vs 3.23 for qwen3.8-mtp:27b (1.16x, 0.87x to 1.47x per prompt).
• Output equals the MTP tag's on 14 of 15 greedy comparisons (the other differs at a near-tie); multi-turn, max_tokens 1 to 16, think on, tools, an image, a 4.5k-token prompt and cancel + retry behave as on qwen3.8-mtp:27b.
• test_model_downloader (catalog / base_model tests) passes, built standalone with g++.

Not in this PR

• src/model_info.json entry and the HF / ModelScope upload of Qwen3.8-27B-DFlash2-NPU2 (needed before flm pull qwen3.8-dflash2:27b works): after the upload, from the HF tree API.
• Windows installer entries: none needed (same qwen3_8mtp_npu.dll).

🤖 Generated with Claude Code

Vendored libraries (IRON PR Check run 41346212324)

c578de45346c36dfcebf45f38fb77a387f9e497aae5fc327ec1a9e894f28c278  src/lib/xrt/libqwen3_8mtp_npu.so
444cbcd24de97a46241ff5ba1e3e98e3809445d35d72ae2805a6e93eff2b4721  src/lib/hrx/libqwen3_8mtp_npu.so
d0ad732e3ba6ce8096d190922c6e25b3db87c5be07783af0433dc35c6f2f053a  src/lib/xrt/qwen3_8mtp_npu.dll
f4b2c9fb6777bc765f744d1964a4c6d1d3c520aaf8772ac53b1d140f6a1422c2  src/lib/xrt/qwen3_8mtp_npu.lib
83f23e72ca00c50d9b457c3996aff8a2da70f5a423f64cc0b7b03194f06e577f  src/lib/hrx/qwen3_8mtp_npu.dll
3ea183f2dbfb8db4a20c966389d959e2eca083121b910d9e4edc3048e24b6969  src/lib/hrx/qwen3_8mtp_npu.lib

Only the Linux XRT library ran locally (HRX and Windows are exercised by this PR's CI). The tag cannot be pulled until the HF / ModelScope upload.