validated intel arc recipes · text, vision, and video
Six validated LLM model families on llama.cpp SYCL and vLLM XPU, plus MiniMax H3 text-to-video on ComfyUI. Exact image digests, required patches, and verifiable execution recipes.
inference cookbook · text & vision
Six model families, one page per family, each pinned to its own container digest and patch list: Qwen3.8-27B, Qwen3.6-35B-A3B, Qwen3.6-27B, Nemotron-3.5-Lightning, Ornith-1.5-35B-A3B, and Muse-Glimmer-30B.
videogen cookbook · video generation
MiniMax H3 FL2VA text-to-video on ComfyUI 0.33.0 with PyTorch 2.13.0+xpu. Six weight files (~40 GiB), a two-card role split (UNet+VAE on GPU0, CLIP on GPU1), and single-card CPU-CLIP fallback.
text & vision recipes
quantized on the B70 with gptqmodel 7.3.2
106.7 tok/s C1 (n=5, cache on) · C5 127.4 Σ-streams
The most-served model on this host. Draft-INT4 overlay lifts the cache-off cell to 112.7; concurrent serving via mixed-split v5 reaches C32 224.2 on short prompts. Also includes Windows 11 standalone kits.
vllm/vllm-openai-xpu@sha256:f01e24f6… · vLLM 0.27.2rc1.dev77SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16 (~18.2 GiB)MTP-preserved community GPTQ, full 128K window
170.9 tok/s MTP4 (p512/g128, n=5) · 101.6 tok/s at 128K
MoE decode reads only ~3 GB of active experts per token, self-limiting to ~140 W. Runs at 150 W cap with identical speed and 80 W less heat. MTP2 beats MTP4 deep in context.
vllm/vllm-openai-xpu@sha256:2c427ef4… · vLLM 0.26.1rc1.dev457llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GPTQ-Int4MTP-preserved community GPTQ, heretic-v2 base
69.3 tok/s MTP4 · 47.6 tok/s held at full 128K
FP16 KV requires ~9.5 GiB at 128K and fails to fit; fp8 halves requirement. In 32K document sessions, follow-up TTFT drops from 38.2 s to 4.1–4.5 s on 91% prefix hits.
2c427ef4 digest · vLLM 0.26.1rc1.dev457llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-GPTQ-Int4GPTQ-INT4 G64 target + NVFP4→BF16 DFlash draft
186.6 tok/s C1 (p2048/g128, n=5, accept 56.5%)
Native MTP accepts 0% on this stack; converted DFlash draft delivers 1.81× speedup over matched control. Capacity run completed 119,904 tokens at 120K limit.
vllm/vllm-openai-xpu@sha256:1da0a954… · vLLM 0.26.1rc1.dev668SergiioB/Nemotron-3.5-Lightning-30B-A3B-GPTQ-Int4-sym-G64-BF16Native-MTP GPTQ-INT4, full 131,072 context
170.8 tok/s MTP4 (p512/g128) · identical to Qwen3.6 MoE
Confirms architectural identity to Qwen3.6-35B-A3B under vLLM XPU MTP speculative decoding. Full 128K context capacity demonstrated on real Battlemage silicon.
2c427ef4 digest · vLLM 0.26.1rc1.dev457SergiioB/Ornith-1.5-35B-A3B-Native-MTP-GPTQ-Int4-sym-G128-BF16Multimodal vision + 128K context on llama.cpp SYCL
42.8 tok/s decode · 3,240 tok/s pp512 · flash attention
Carries GGUF workflows on SYCL. Flash attention with q8_0 key cache and q4_1 value cache delivers high-speed multimodal inference at 128K context window.
llama.cpp SYCL (DPC++ 2024.2.1) with -DGGML_SYCL_F16=ONMuse-Glimmer-30B-Q4_K_M.gguf-ctk q8_0 -ctv q4_1 -fa 1 (flash attention mandatory)