--- license: apache-2.0 base_model: Qwen/Qwen3.5-122B-A10B base_model_relation: quantized library_name: gguf pipeline_tag: text-generation tags: - gguf - qwen3.5 - qwen3.5-moe - rocmfp4 - imatrix - amd - vulkan - rocm - mtp - speculative-decoding - experimental ---
# Qwen3.5 122B-A10B · ROCmFP4 iMatrix ### The official Qwen checkpoint in a compact, importance-calibrated ROCmFP4 GGUF **122B total · 10B active · 60.70 GiB · 28.50 tok/s MTP-off · BF16 KLD 0.041366 · Decode 28.505** Decode speed + 36.89% faster Size - 13.47gb smaller
> [!IMPORTANT] > This GGUF uses custom ROCmFP4 tensor types. It requires > [ROCmFPX](https://github.com/charlie12345/ROCmFPX) or a runtime with equivalent > support. Stock `llama.cpp`, Ollama, LM Studio, and similar stock runtimes > cannot load it. I also suggest usign the template from https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates Froggeric is tha dude. This release was tested only on AMD Strix Halo / `gfx1151`. Other AMD targets may require a different ROCmFPX build and are untested; CPU-only, NVIDIA, and non-ROCmFPX runtimes are not supported by this card. ## Downloads | Artifact | Direct download | |---|---| | **Main iMatrix model — 60.70 GiB** | **[Download `Qwen3.5-122B-A10B-ROCmFP4-iMatrix.gguf`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-ROCmFP4-iMatrix.gguf?download=true)** | | **Optional MTP companion — 2.14 GiB** | **[Download `Qwen3.5-122B-A10B-ROCmFP4-MTP.gguf`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-ROCmFP4-MTP.gguf?download=true)** | | Calibration matrix — 342.28 MiB | [Download `Qwen3.5-122B-A10B-Bartowski.imatrix`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-Bartowski.imatrix?download=true) | [Browse every repository file →](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/tree/main) ## What it is This is an iMatrix-calibrated quantization of the official [`Qwen/Qwen3.5-122B-A10B`](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) checkpoint. It was built with ROCmFPX's compact `Q4_0_ROCMFP4_STRIX_LEAN` recipe; the hardware-oriented preset name is left out of the public filename. | | Result | |---|---:| | **BF16 mean KLD** | **0.041366 ± 0.002531** | | **Greedy decode** | **28.505 tok/s** | | **Sampled decode** | **28.485 tok/s** | | **4,277-token prefill** | **356.900 tok/s** | ## Files | File | Purpose | Size | SHA-256 | |---|---|---:|---| | [`Qwen3.5-122B-A10B-ROCmFP4-iMatrix.gguf`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-ROCmFP4-iMatrix.gguf?download=true) | Main text model | 65,184,265,120 bytes | `9f44eb8a8693f46af6e1b06f6219229eb074c1ec5527798e8a18d68034b381c8` | | [`Qwen3.5-122B-A10B-ROCmFP4-MTP.gguf`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-ROCmFP4-MTP.gguf?download=true) | Optional external MTP companion | 2,294,290,272 bytes | `f59efaa7c184042a940df322d81921254b8e178d6eefc0fd5cdd7c5b0a9acbe3` | | [`Qwen3.5-122B-A10B-Bartowski.imatrix`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/Qwen3.5-122B-A10B-Bartowski.imatrix?download=true) | Calibration matrix | 358,906,272 bytes | `e8bfa39dd663e70655035ad53bf715069b7f55175a8877a36bbbe18a0131fed6` | | [`chat_template.jinja`](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-ROCmFP4-iMatrix-GGUF/resolve/main/chat_template.jinja?download=true) | Pinned Qwen3.5 chat template | 7,756 bytes | `a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715` | The main model is complete and runs independently. The 23-tensor, one-layer MTP file is an optional external companion; no MTP layer is embedded in the main GGUF. Runtimes call this role a *draft model*, but `Draft` is omitted from the public filename because it does not mean unfinished. ## Quality: stock and Heretic KLD This is the card's only comparison between the stock and Heretic builds. Each ROCmFP4 iMatrix model was replayed against saved distributions from its own exact BF16 parent, using the same runtime, WikiText-2 token sequence, ROCm0 backend, two 512-token chunks, and 510 evaluated next-token distributions. | Build | Mean KLD from its BF16 parent | |---|---:| | **Stock iMatrix** | **0.041366 ± 0.002531** | | [Heretic iMatrix](https://huggingface.co/vmlinux/Qwen3.5-122B-A10B-Heretic-ROCmFP4-iMatrix-GGUF) | 0.041395 ± 0.002697 | Lower KLD means the quantized model's next-token distribution stayed closer to its own BF16 source. It does **not** rank the underlying stock and Heretic models, and it is not an intelligence or benchmark score. The two-chunk scope is deliberately bounded: this is the demonstrated-safe BF16 workload on the tested 128 GB system. Treat it as a controlled calibration comparison, not a completed large-corpus acceptance gate. The Heretic card's headline uses a separate 11-chunk replay; its value is therefore not expected to match the two-chunk, method-matched figure reported here. ## Performance Measured on a 128 GB AMD Strix Halo system with Vulkan/RADV, 131,072 context, parallel 1, batch/ubatch 2048/1024, Q8_0 KV cache, flash attention, and MTP off. Values are medians from repeated runs of this stock artifact. | Workload | Prompt tokens | Repeats | Decode | Prompt processing | |---|---:|---:|---:|---:| | Greedy, 256 generated tokens | 52 | 5 | **28.505 tok/s** | 71.934 tok/s | | Sampled, 256 generated tokens | 52 | 3 | **28.485 tok/s** | 70.957 tok/s | | 4,277-token prefill + 128 generated | 4,277 | 3 | **28.107 tok/s** | **356.900 tok/s** | These are single-system measurements, not general performance guarantees. The MTP companion is included for compatible runtimes, but this card does not claim an MTP speed result for the iMatrix artifact. ## Quantization and matrix provenance The model was quantized once from the validated BF16 GGUF using the `Q4_0_ROCMFP4_STRIX_LEAN` preset and the importance matrix published with [`bartowski/Qwen_Qwen3.5-122B-A10B-GGUF`](https://huggingface.co/bartowski/Qwen_Qwen3.5-122B-A10B-GGUF) at revision `f89fb67573c0155d8e5b6556204d86c75cdce0d8`. | Tensor type | Count | |---|---:| | `Q4_0_ROCMFP4_FAST` | 457 | | `Q4_0_ROCMFP4` | 60 | | `F32` | 361 | | `Q5_K` | 1 | In this preset, attention K/V tensors retain dual-scale ROCmFP4 protection, most transformer weights use the compact FAST layout, and token embeddings use Q5_K while the separate output head uses the FAST layout. The artifact contains 879 tensors across 48 blocks and no embedded MTP layer. The matrix contains 612 entries from 802 × 512-token chunks. Structural inspection found finite paired tensors and 37,323 of 37,332 expert count slots covered; nine slots had zero observations. A quantizer dry run accepted all 612 importance entries. The matrix publisher identified the official upstrQwen3-235B-A22Beam model but did not pin an immutable upstream weight revision, so that provenance limitation is recorded here rather than silently inferred. ## Run Use the ROCmFPX-built `llama-server`, not a stock `llama.cpp` binary: ```bash llama-server \ --model Qwen3.5-122B-A10B-ROCmFP4-iMatrix.gguf \ --host 127.0.0.1 --port 8080 \ -dev Vulkan0 --n-gpu-layers 999 \ --ctx-size 131072 --parallel 1 \ --flash-attn on --batch-size 2048 --ubatch-size 1024 \ --cache-type-k q8_0 --cache-type-v q8_0 \ --jinja --reasoning-format deepseek ``` Adjust context and cache settings for your memory budget. The tested runtime was [charlie12345/ROCmFPX](https://github.com/charlie12345/ROCmFPX) commit `a6a93765f7ce9779c13f9881164a65f7a9f31198`, built in Release mode for `gfx1151` with Vulkan and HIP enabled. Correct inference also requires the duplicate Qwen3.5 MoE down-scale fix from [`llama.cpp` PR #24331](https://github.com/ggml-org/llama.cpp/pull/24331), commit [`02810c7`](https://github.com/ggml-org/llama.cpp/commit/02810c7aa89b8100b90b7b0f5e96bc55aafd3d0a): without it, the expert down scale is applied twice. The pinned ROCmFPX commit did not yet contain the fix, so the tested runtime applied that exact nine-line correction locally; use a newer ROCmFPX revision or equivalent runtime that includes it. Results with other revisions have not yet been established. The GGUF embeds the same chat template shipped separately in this repository, so the command uses `--jinja` without an external template path. `--reasoning-format deepseek` is the tested llama.cpp parser for this template's reasoning output, not a claim that the model is a DeepSeek derivative. The model file alone occupies 60.70 GiB; leave additional memory for the runtime and KV cache, especially at the tested 131,072-token context. To enable the optional companion, append: ```bash --spec-type draft-mtp \ --spec-draft-model Qwen3.5-122B-A10B-ROCmFP4-MTP.gguf \ --spec-draft-device Vulkan0 --spec-draft-ngl 999 \ --spec-draft-type-k f16 --spec-draft-type-v f16 \ --spec-draft-n-max 2 --spec-draft-p-min 0.6 \ --spec-draft-p-split 0.10 --spec-draft-backend-sampling ``` ## Tested system | | | |---|---| | **Platform** | AMD Strix Halo, 128 GB unified memory | | **GPU target** | Radeon 8060S / `gfx1151` | | **Backend** | Vulkan / RADV for serving; ROCm for bounded BF16 KLD collection | | **Kernel** | Linux 6.17.0-1028-oem | | **Mesa** | 25.2.8 | No vision projector is included; treat this release as text-only. ## Lineage and credits - **Official model and MTP tensors:** [Qwen/Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B), revision `dc4d348443bc740c68e2d77492492c11606384d5`, Apache 2.0. - **Importance matrix:** [bartowski/Qwen_Qwen3.5-122B-A10B-GGUF](https://huggingface.co/bartowski/Qwen_Qwen3.5-122B-A10B-GGUF), revision `f89fb67573c0155d8e5b6556204d86c75cdce0d8`. - **ROCmFP4 implementation, quantizer, and compatible runtime:** [charlie12345/ROCmFPX](https://github.com/charlie12345/ROCmFPX). - **Strix Halo deployment stack:** [hec-ovi/llama-vulkan-strix](https://github.com/hec-ovi/llama-vulkan-strix). - **Conversion, validation, benchmarking, and packaging:** `vmlinux`. Please preserve this lineage, the Apache 2.0 license, and a description of your changes when redistributing derivatives. ## License The distributed model derivative is provided under the upstream Apache License 2.0, whose text is included as `LICENSE`. Runtime and tooling repositories retain their own licenses. No runtime source code is bundled in this model repository.