#6317 [Engine] Scope for an opt-in Vulkan research path on the Intel Arc B390 iGPU
open · @grigio · 3 评论 · 去 GitHub 看
BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindowsLinux
说明
### Area [Engine] (speed, kernels, models, the server and API) ## What do you want to do? Agree the scope for a narrow, **opt-in Vulkan research path** that serves one Qwen3.8-35B-A3B GGUF on an Intel Arc B390 iGPU (Panther Lake) from a Linux source checkout. It would not change the default path: `strata launch --backend vulkan` builds with `STRATA_ENABLE_VULKAN=ON` (OFF by default), uses Vulkan + C++17 only, and leaves the CUDA/HIP/SYCL/Metal flow untouched. A native engine streams the model's MoE experts from the GGUF into a bounded slot pool; an optional `--vulkan-impl bridge` drives llama.cpp's Vulkan backend. It is a research preview: no released engine package, Linux only, one model family, one token at a time, no prompt batching, no VRAM planner or drafter. I have a branch prepared locally (`vulkan_arc`) but will **not** open a PR unless this scope is agreed here first, as CONTRIBUTING.md requires for substantial features. ## Why this is not the declined general Vulkan backend - #1214 was closed as not planned: *"CUDA, HIP and SYCL already cover NVIDIA, AMD and Intel cards, and a Vulkan backend would be a separate engine."* This proposal is not a general multi-vendor Vulkan backend. It targets one documented gap: iGPUs that share system RAM. - #515: the maintainer wrote that for the B390 *"Windows has no setup path for the Intel engine, and the planning for a GPU that shares system RAM does not exist yet (the port assumes a discrete card with its own VRAM)."* - #2412 (B390 on the SYCL port): the TTM page pool hides about 30 GB from `MemAvailable`. That issue was closed in the 0.2.0-beta tracker reset, whose comment says *"Closed here does not mean rejected"* and invites `/reopen`. I am not treating that closure as a rejection. - #943 (an external engine adapter) was declined as out of scope: *"it would work well as an external plugin."* If this work belongs outside the repository for the same reason, please say so here and I will not open the PR. The ask is a maintainer decision: does this narrow, opt-in iGPU path belong in-tree as a research preview, or should it stay external? If the #1214 decision has changed with the new B390 evidence, CONTRIBUTING.md asks that it be brought back to the original discussion — this issue is that discussion. ## What do you do today instead? The SYCL port (#423, docs/INTEL_ARC.md) runs on the B390 on Linux with oneAPI, but it has no shared-memory-aware planner and #2412 reports long outputs degenerating. There is no setup path on Windows. The prepared path builds without oneAPI or Level Zero and needs only a Vulkan 1.3 driver plus the GGUF. ## Validation done for this scope request (separate from what was not run) Passed just now on this checkout: - Fresh `cmake -S engine -B <tmp> -G Ninja -DSTRATA_ENABLE_VULKAN=ON` configures and builds all 7 test executables; 13 shaders compile with glslangValidator 16.6.0. - All 6 Vulkan kernel test binaries pass (router, combine, bf16, sampler, q8, iq), 0 failures each. - `engine/vulkan/asan_check.sh` passes all 7 binaries, including the GGUF-backed stream test, with no sanitizer reports. - The checked-in `.spv` files reproduce byte-for-byte from their `.comp` sources. Not re-run during this check: decode-speed A/B, the 30-minute soak, NVMe throughput, and the end-to-end model comparison. The prepared `docs/VULKAN.md` records those from runs on a Core Ultra X7 358H / Arc B390 with about 30 GB of shared RAM, Mesa ANV Vulkan 1.4 and a YMTC PC411 NVMe: native about 5.5 tok/s versus about 1 tok/s with the earlier 64-slot cache, bridge about 31 tok/s, teacher-forcing agreement on 214 positions, and 555 requests / 0 bad tokens in 30 minutes. Those are single-machine numbers, not a claim about other hardware. ## Tradeoffs and limits - About 15k lines across 119 files: a new engine directory, CMake option, CLI and doctor integration, docs and tests. One feature, one PR, but I am happy to split it if you prefer. - Opt-in only: CMake option OFF, `--backend vulkan` explicit, default path unchanged. - Linux and Vulkan 1.3 only; one GGUF; no Windows or macOS; no released installer package. - Native engine is about 5.5 tok/s; the faster bridge does not stream experts. Neither matches the maintained SYCL/CUDA paths.
本站相关内容
相关页面的快捷入口。