{ "schema_version": "1.0", "database_version": "2026.11.06-2", "last_reviewed": "2026-11-06", "scope": "Publicly documented accelerator software compatibility and integration evidence. Entries are not certification or production-readiness claims. Product-specific links are only added when the cited source explicitly names that product or generation; otherwise the record remains platform-level.", "status_definitions": { "upstream_supported": "The upstream project documentation lists the hardware platform as supported.", "vendor_plugin_supported": "A vendor or project-maintained out-of-tree plugin documents a released integration and its scope.", "vendor_framework": "A vendor-maintained framework backend or SDK documents the integration and version matrix.", "experimental": "A development branch, preview, or upgrade candidate exists, but its own documentation does not yet mark it as a supported release.", "in_progress": "The upstream project explicitly marks the integration as in progress.", "needs_review": "Public documentation contains conflicting or incomplete version mapping; do not use as a validated compatibility claim.", "vendor_reported": "A vendor has publicly reported a model port or validation case; independent reproduction is not established." }, "platforms": [ { "id": "nvidia-gpu", "vendor": "NVIDIA", "name": "NVIDIA GPU", "hardware_scope": [ "NVIDIA CUDA GPUs; exact SKU and architecture support is release-specific" ], "vendor_stack": [ "CUDA", "TensorRT", "NCCL" ], "evidence_urls": [ "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "amd-gpu", "vendor": "AMD", "name": "AMD GPU", "hardware_scope": [ "Instinct MI300X / MI325X / MI350X / MI355X as listed by SGLang; verify each engine release" ], "vendor_stack": [ "ROCm", "HIP", "RCCL" ], "evidence_urls": [ "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "intel-xpu", "vendor": "Intel", "name": "Intel GPU / CPU", "hardware_scope": [ "Arc and Arc Pro B-Series GPUs; Xeon CPUs as listed by SGLang" ], "vendor_stack": [ "oneAPI", "OpenVINO", "oneDNN" ], "evidence_urls": [ "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "huawei-ascend", "vendor": "Huawei", "name": "Ascend NPU", "hardware_scope": [ "Atlas A2 / A3 systems; Atlas 300I Duo is marked experimental in vLLM-Ascend prerequisites", "SGLang README separately lists A2 / A3 / 950PR / DT" ], "vendor_stack": [ "CANN", "torch-npu", "HCCL" ], "evidence_urls": [ "https://github.com/vllm-project/vllm-ascend/blob/main/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "moore-threads", "vendor": "Moore Threads", "name": "MUSA GPU", "hardware_scope": [ "MTT S5000" ], "vendor_stack": [ "MUSA", "torch_musa", "MATE", "torchada" ], "evidence_urls": [ "https://github.com/MooreThreads/vllm-musa/blob/v0.32.0-dev/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "metax", "vendor": "MetaX", "name": "MetaX GPU", "hardware_scope": [ "MetaX C-series as listed by the vLLM-MetaX plugin" ], "vendor_stack": [ "MACA", "MXMACA" ], "evidence_urls": [ "https://github.com/MetaX-MACA/vLLM-metax/blob/master/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "hygon", "vendor": "Hygon", "name": "Hygon HCU / DCU", "hardware_scope": [ "Hygon HCU / DCU; SGLang marks integration in progress" ], "vendor_stack": [ "DTK", "torch-dcu" ], "evidence_urls": [ "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "cambricon", "vendor": "Cambricon", "name": "Cambricon MLU", "hardware_scope": [ "Cambricon MLU; SGLang marks integration in progress" ], "vendor_stack": [ "Neuware", "CNToolkit", "torch_mlu" ], "evidence_urls": [ "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "iluvatar", "vendor": "Iluvatar CoreX", "name": "Iluvatar CoreX GPU", "hardware_scope": [ "CoreX GPUs; plugin build examples identify ivcore11 and ivcore30 architectures" ], "vendor_stack": [ "CoreX runtime", "IXUCA SDK", "ixRT", "IGIE", "ixformer" ], "evidence_urls": [ "https://github.com/Deep-Spark/vllm-iluvatar/blob/release/v0.27.1/README.md", "https://github.com/Deep-Spark/DeepSparkInference/blob/master/README.md", "https://github.com/sgl-project/sglang/blob/main/README.md" ] }, { "id": "enflame", "vendor": "Enflame", "name": "Enflame GCU", "hardware_scope": [ "CloudBlazer Inference S60 as listed in torch-gcu README" ], "vendor_stack": [ "TopsRider", "TopsRuntime", "TopsAten", "ECCL", "torch-gcu" ], "evidence_urls": [ "https://github.com/EnflameTechnology/torch-gcu/blob/main/README.md" ] }, { "id": "biren", "vendor": "Biren", "name": "Biren GPU", "hardware_scope": [ "BR166 for the vendor-reported DeepSeek-V4 preview adaptation example" ], "vendor_stack": [ "BIRENSUPA", "br_pytorch", "AIModelMaster" ], "evidence_urls": [ "https://www.birentech.com/news/lo81mr3idb49bsdjf8hdd2bb/", "https://www.birentech.com/news/otr0bf5evzdwwivrjt6h3322/" ] }, { "id": "intel-gaudi", "vendor": "Intel", "name": "Intel Gaudi", "hardware_scope": [ "Intel Gaudi 2 and Gaudi 3 accelerators; exact SKU, board form factor and software release support are version-specific." ], "vendor_stack": [ "Intel Gaudi Software", "SynapseAI", "PyTorch bridge", "HCCL", "vLLM Hardware Plugin for Intel Gaudi" ], "evidence_urls": [ "https://docs.habana.ai/en/latest/", "https://docs.habana.ai/en/latest/Release_Notes/GAUDI_Release_Notes.html" ] }, { "id": "google-tpu", "vendor": "Google", "name": "Google Cloud TPU", "hardware_scope": [ "TPU v6e (Trillium) and TPU7x (Ironwood); runtime and framework support differ by TPU generation." ], "vendor_stack": [ "JAX", "XLA", "TorchTPU", "vLLM-TorchTPU", "SGLang-TorchTPU", "GKE / XPK" ], "evidence_urls": [ "https://docs.cloud.google.com/tpu/docs/pytorch-on-tpu", "https://docs.cloud.google.com/tpu/docs/tpu-inference", "https://docs.cloud.google.com/tpu/docs/runtimes" ] }, { "id": "aws-trainium", "vendor": "Amazon Web Services", "name": "AWS Trainium", "hardware_scope": [ "Trainium2 (Trn2) and Trainium3 (Trn3) instances; Neuron SDK and model support are release-specific." ], "vendor_stack": [ "AWS Neuron SDK", "TorchNeuron", "vLLM Neuron", "Neuron Kernel Interface (NKI)", "JAX Neuron plugin" ], "evidence_urls": [ "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/index.html", "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/inference.html" ] }, { "id": "microsoft-maia", "vendor": "Microsoft", "name": "Microsoft Azure Maia", "hardware_scope": [ "Azure Maia 100 and Maia 200 accelerators in Microsoft-operated Azure infrastructure." ], "vendor_stack": [ "Azure infrastructure software", "Microsoft internal compiler/runtime stack; public developer compatibility details remain limited." ], "evidence_urls": [ "https://techcommunity.microsoft.com/blog/azureinfrastructureblog/inside-maia-100-revolutionizing-ai-workloads-with-microsofts-custom-ai-accelerat/4229118", "https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/" ] }, { "id": "meta-mtia", "vendor": "Meta", "name": "Meta MTIA", "hardware_scope": [ "Meta MTIA v2 and MTIA 300 internal data-center accelerators." ], "vendor_stack": [ "Meta internal PyTorch / MTIA stack", "Public external SDK and compatibility matrix not identified in the reviewed sources." ], "evidence_urls": [ "https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/", "https://engineering.fb.com/2026/08/24/networking-traffic/mtia-300-meta-training-chip-built-in-nics/" ] }, { "id": "cerebras-wse", "vendor": "Cerebras", "name": "Cerebras Wafer-Scale Engine", "hardware_scope": [ "Cerebras WSE-2 and WSE-3 wafer-scale systems; exact deployment SKU and service availability are configuration-specific." ], "vendor_stack": [ "Cerebras PyTorch", "Cerebras ModelZoo", "Cerebras SDK", "Cerebras Inference SDK" ], "evidence_urls": [ "https://docs.cerebras.ai/", "https://github.com/Cerebras/modelzoo" ] } ], "integrations": [ { "id": "nvidia-vllm", "platform_id": "nvidia-gpu", "layer": "serving", "project": "vLLM", "project_url": "https://github.com/vllm-project/vllm", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "NVIDIA CUDA GPU path; exact model, GPU generation, CUDA and kernel coverage remain release-specific.", "versions": { "framework_release": "v0.33.0rc3 (release candidate; latest release checked 2026-11-05)" }, "limitations": "The upstream installation guide lists NVIDIA CUDA as a core hardware path. This entry does not imply every model or feature works on every GPU.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "checked_at": "2026-11-05", "claim": "Upstream installation documentation lists NVIDIA CUDA as a supported hardware platform." } ] }, { "id": "nvidia-sglang", "platform_id": "nvidia-gpu", "layer": "serving", "project": "SGLang", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "NVIDIA GPU families listed in the current upstream hardware table; verify exact SKU and feature coverage.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); hardware list read from current main README" }, "limitations": "The upstream hardware table is broader than any single tested model / feature combination.", "last_checked": "2026-11-06", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-06", "claim": "NVIDIA is listed in SGLang Supported Hardware." } ], "product_ids": [ "b200-sxm", "gb200-nvl72" ] }, { "id": "amd-vllm", "platform_id": "amd-gpu", "layer": "serving", "project": "vLLM ROCm", "project_url": "https://github.com/vllm-project/vllm", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "AMD ROCm hardware path; exact GPU generation, ROCm version and feature support must be checked against the release docs.", "versions": { "framework_release": "v0.33.0rc3 (release candidate; latest release checked 2026-11-05)" }, "limitations": "Core platform support is not a guarantee of feature parity with CUDA.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "checked_at": "2026-11-05", "claim": "Upstream installation documentation lists AMD ROCm as a supported hardware platform." }, { "type": "upstream_docs", "url": "https://docs.vllm.ai/en/latest/features/speculative_decoding/fly/", "checked_at": "2026-11-06", "claim": "The vLLM FLY verification matrix explicitly lists AMD MI355X with Llama 3.1 405B Instruct FP8-KV and Qwen3-235B-A22B-Thinking-2507-FP8." } ], "product_ids": [ "mi355x" ] }, { "id": "amd-sglang", "platform_id": "amd-gpu", "layer": "serving", "project": "SGLang", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "Instinct MI300X, MI325X, MI350X and MI355X as listed by SGLang.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); hardware list read from current main README" }, "limitations": "Validate exact ROCm release, GPU generation, attention backend and model support.", "last_checked": "2026-11-06", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-06", "claim": "SGLang Supported Hardware table lists MI300X / MI325X / MI350X / MI355X." } ], "product_ids": [ "mi350x", "mi355x" ] }, { "id": "intel-vllm", "platform_id": "intel-xpu", "layer": "serving", "project": "vLLM Intel XPU", "project_url": "https://github.com/vllm-project/vllm", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "Intel XPU hardware path; exact GPU model and feature coverage are release-specific.", "versions": { "framework_release": "v0.33.0rc3 (release candidate; latest release checked 2026-11-05)" }, "limitations": "Do not conflate Intel XPU GPU support with Intel CPU support; they have separate execution paths.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/README.md", "checked_at": "2026-11-05", "claim": "Upstream installation documentation lists Intel XPU as a supported GPU platform." } ] }, { "id": "intel-sglang", "platform_id": "intel-xpu", "layer": "serving", "project": "SGLang Intel GPU / CPU", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "Arc / Arc Pro B-Series GPUs and Xeon CPUs as listed by SGLang.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); hardware list read from current main README" }, "limitations": "GPU and CPU support are separate paths; validate model, dtype and operator coverage independently.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "SGLang hardware table lists Intel Arc / Arc Pro B-Series GPUs and Xeon CPUs." } ] }, { "id": "ascend-vllm", "platform_id": "huawei-ascend", "layer": "serving", "project": "vLLM Ascend", "project_url": "https://github.com/vllm-project/vllm-ascend", "maintainer_type": "project_community", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Atlas A2 / A3 inference and training systems; Atlas 300I Duo is explicitly experimental in prerequisites.", "versions": { "plugin_release": "v0.30.0rc1 recommended quick-start release", "vllm": "same version as vllm-ascend", "cann": "9.1.0", "pytorch": "2.10.0", "torch_npu": "2.10.0.post4" }, "limitations": "The project describes itself as a community-maintained plugin. Use the plugin's own support matrix and versioning policy; do not mix arbitrary upstream vLLM versions.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/vllm-project/vllm-ascend/blob/main/README.md", "checked_at": "2026-11-05", "claim": "README lists prerequisites and recommended v0.30.0rc1 quick-start, and describes the project as a community-maintained hardware plugin." }, { "type": "upstream_docs", "url": "https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/", "checked_at": "2026-11-05", "claim": "Project-maintained support matrix for models and features." } ] }, { "id": "ascend-sglang", "platform_id": "huawei-ascend", "layer": "serving", "project": "SGLang Ascend", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "A2, A3 and 950PR / DT NPUs as listed in SGLang README.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); hardware list read from current main README" }, "limitations": "SGLang's platform-level support does not imply feature parity across all Ascend generations.", "last_checked": "2026-11-06", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-06", "claim": "Huawei Ascend row lists A2, A3, 950PR / DT." } ], "product_ids": [ "ascend-950pr", "ascend-950dt" ] }, { "id": "mthreads-vllm", "platform_id": "moore-threads", "layer": "serving", "project": "vLLM-MUSA", "project_url": "https://github.com/MooreThreads/vllm-musa", "maintainer_type": "vendor_plugin", "status": "experimental", "confidence": "high", "hardware_scope": "MUSA GPUs; current README includes S5000 model recipes.", "versions": { "development_branch": "v0.32.0-dev", "pytorch": "2.11.0.post2", "torch_musa": "2.11.0.post2", "musa": "5.2.0", "mate": "0.2.8.post1", "torchada": "0.1.105", "release_image": "v0.32.0 (published image; verify exact package pins separately)" }, "limitations": "README explicitly calls the v0.32.0-dev branch an upgrade candidate and says promotion requires build and model-smoke validation. Do not treat the development branch as a validated supported release.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/MooreThreads/vllm-musa/blob/v0.32.0-dev/README.md", "checked_at": "2026-11-05", "claim": "README documents the pinned stack and labels the development branch an upgrade candidate." } ] }, { "id": "mthreads-sglang", "platform_id": "moore-threads", "layer": "serving", "project": "SGLang Moore Threads", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "upstream_supported", "confidence": "high", "hardware_scope": "MTT S5000 GPUs as listed by SGLang.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); hardware list read from current main README" }, "limitations": "Validate exact MUSA software stack and model / kernel coverage.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "Moore Threads row lists MTT S5000 in Supported Hardware." } ] }, { "id": "metax-vllm", "platform_id": "metax", "layer": "serving", "project": "vLLM-MetaX", "project_url": "https://github.com/MetaX-MACA/vLLM-metax", "maintainer_type": "vendor_plugin", "status": "needs_review", "confidence": "medium", "hardware_scope": "MetaX C-series; deployment README currently points to MetaX developer-community Docker images.", "versions": { "plugin_release": "v0.28.0 (latest release note checked 2026-11-05)", "framework_version_claim": "vLLM v0.28.0 in release notes", "python": "3.10–3.12", "backend": "MACA" }, "limitations": "The README's release notes say plugin v0.28.0 aligns with upstream vLLM v0.28.0, but its maintained-branch table maps releases/v0.28.0 to vLLM v0.26.0. This conflict must be resolved before treating the version pair as validated.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/MetaX-MACA/vLLM-metax/blob/master/README.md", "checked_at": "2026-11-05", "claim": "README describes the plugin, C-series hardware, Docker-only deployment, release notes and a conflicting branch compatibility table." } ] }, { "id": "metax-sglang", "platform_id": "metax", "layer": "serving", "project": "SGLang MetaX integration", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "in_progress", "confidence": "high", "hardware_scope": "MetaX integration; no generally supported device family is claimed by the SGLang README.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); integration status read from current main README" }, "limitations": "SGLang explicitly lists MetaX under integrations in progress.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "MetaX is listed under integrations in progress." } ] }, { "id": "hygon-sglang", "platform_id": "hygon", "layer": "serving", "project": "SGLang Hygon HCU / DCU integration", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "in_progress", "confidence": "high", "hardware_scope": "Hygon HCU / DCU integration.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); integration status read from current main README" }, "limitations": "SGLang explicitly lists Hygon HCU / DCU under integrations in progress.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "Hygon HCU / DCU is listed under integrations in progress." } ] }, { "id": "cambricon-sglang", "platform_id": "cambricon", "layer": "serving", "project": "SGLang Cambricon MLU integration", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "in_progress", "confidence": "high", "hardware_scope": "Cambricon MLU integration.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); integration status read from current main README" }, "limitations": "SGLang explicitly lists Cambricon MLU under integrations in progress.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "Cambricon MLU is listed under integrations in progress." } ] }, { "id": "iluvatar-vllm", "platform_id": "iluvatar", "layer": "serving", "project": "vLLM-Iluvatar", "project_url": "https://github.com/Deep-Spark/vllm-iluvatar", "maintainer_type": "vendor_plugin", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Iluvatar CoreX GPUs; build instructions give ivcore11 and ivcore30 as example architecture targets.", "versions": { "plugin_branch": "release/v0.27.1", "vllm": ">=0.27.1,<0.28.0", "corex_runtime_driver": ">=5.0.0", "corex_pytorch": ">=2.10.0", "ixformer": ">=0.7.0", "python": ">=3.10" }, "limitations": "The plugin pins an older vLLM range and CoreX-specific packages are supplied separately. Compatibility with vLLM 0.33.0rc3 is not established.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/Deep-Spark/vllm-iluvatar/blob/release/v0.27.1/README.md", "checked_at": "2026-11-05", "claim": "README lists exact vLLM, CoreX runtime, CoreX PyTorch, ixformer and Python requirements." } ] }, { "id": "iluvatar-sglang", "platform_id": "iluvatar", "layer": "serving", "project": "SGLang Iluvatar CoreX integration", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "in_progress", "confidence": "high", "hardware_scope": "Iluvatar CoreX integration.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-05); integration status read from current main README" }, "limitations": "SGLang explicitly lists Iluvatar CoreX under integrations in progress.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-05", "claim": "Iluvatar CoreX is listed under integrations in progress." } ] }, { "id": "enflame-torch-gcu", "platform_id": "enflame", "layer": "framework", "project": "torch-gcu", "project_url": "https://github.com/EnflameTechnology/torch-gcu", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "CloudBlazer Inference S60, the hardware model named in the public README.", "versions": { "torch_gcu": "2.10.0", "pytorch": "2.10.0", "topsrider": "v3.7.1", "python": "3.9, 3.10, 3.12" }, "limitations": "This is a PyTorch backend integration, not evidence of vLLM or SGLang production support. The README documents torch.compile / Inductor, AMP, profiler, GCU Graph, ECCL and CUDA migration.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/EnflameTechnology/torch-gcu/blob/main/README.md", "checked_at": "2026-11-05", "claim": "README lists the PyTorch / Python matrix, TopsRider compatibility and S60 hardware support." } ] }, { "id": "biren-deepseek-v4-port", "platform_id": "biren", "layer": "model_porting", "project": "DeepSeek-V4 preview adaptation on BR166", "project_url": "https://www.birentech.com/news/lo81mr3idb49bsdjf8hdd2bb/", "maintainer_type": "vendor", "status": "vendor_reported", "confidence": "medium", "hardware_scope": "Biren BR166 GPU.", "versions": { "model": "DeepSeek-V4 preview", "reported_date": "2026-04-28" }, "limitations": "Vendor-reported adaptation / inference validation. Public evidence in this review does not establish an independently maintained upstream vLLM or SGLang backend, full model coverage or reproducible benchmark data.", "last_checked": "2026-11-05", "evidence": [ { "type": "vendor_announcement", "url": "https://www.birentech.com/news/lo81mr3idb49bsdjf8hdd2bb/", "checked_at": "2026-11-05", "claim": "Biren reported DeepSeek-V4 preview adaptation and inference validation on BR166." }, { "type": "vendor_announcement", "url": "https://www.birentech.com/news/otr0bf5evzdwwivrjt6h3322/", "checked_at": "2026-11-05", "claim": "A separate vendor case describes SenseNova U1 adaptation using br_pytorch, BIRENSUPA and AIModelMaster." } ] }, { "id": "intel-gaudi-pytorch", "platform_id": "intel-gaudi", "layer": "framework", "project": "Intel Gaudi PyTorch bridge", "project_url": "https://docs.habana.ai/en/latest/PyTorch/index.html", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Gaudi 2 and Gaudi 3; exact operator and model support is tied to the versioned Gaudi support matrix.", "versions": { "gaudi_software": "1.24.2", "execution_mode": "Eager + torch.compile recommended; Lazy mode legacy" }, "limitations": "Vendor-provided PyTorch integration. Check the versioned support matrix and known limitations for the exact model, operator, OS and accelerator. Current documentation recommends Eager mode with torch.compile; Lazy mode is a legacy fallback.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://docs.habana.ai/en/latest/PyTorch/index.html", "checked_at": "2026-11-05", "claim": "Intel Gaudi documentation describes the PyTorch bridge, training and inference paths for Gaudi accelerators." }, { "type": "upstream_docs", "url": "https://docs.habana.ai/en/latest/PyTorch/Reference/PyTorch_Gaudi_Theory_of_Operations.html", "checked_at": "2026-11-05", "claim": "The current Gaudi PyTorch documentation describes Eager mode with torch.compile as recommended and Lazy mode as legacy." } ], "product_ids": [ "gaudi2", "gaudi3" ] }, { "id": "intel-gaudi-vllm", "platform_id": "intel-gaudi", "layer": "serving", "project": "vLLM Hardware Plugin for Intel Gaudi", "project_url": "https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Model_Serving.html", "maintainer_type": "vendor_plugin", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Gaudi 2 and Gaudi 3 subject to the matching Gaudi software and plugin release.", "versions": { "gaudi_software": "1.24.2", "vllm_plugin": "0.30.0", "vllm_base": "0.30.0" }, "limitations": "Vendor-maintained hardware plugin. Model, quantization, kernel and feature coverage are release-specific; plugin support is not equivalent to every upstream vLLM feature working on every Gaudi SKU.", "last_checked": "2026-11-05", "evidence": [ { "type": "release_note", "url": "https://docs.habana.ai/en/latest/Release_Notes/GAUDI_Release_Notes.html", "checked_at": "2026-11-05", "claim": "Gaudi 1.24.2 release notes identify compatibility with vLLM Hardware Plugin 0.30.0 based on upstream vLLM 0.30.0." } ], "product_ids": [ "gaudi2", "gaudi3" ] }, { "id": "google-tpu-torchtpu", "platform_id": "google-tpu", "layer": "framework", "project": "TorchTPU", "project_url": "https://docs.cloud.google.com/tpu/docs/pytorch-on-tpu", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Cloud TPU v6e and TPU7x with generation-specific runtime images and software versions.", "versions": { "framework": "PyTorch", "backend": "TorchTPU" }, "limitations": "TorchTPU is the current PyTorch-native TPU backend described by Google Cloud. Do not treat legacy PyTorch/XLA instructions as the recommended path for new workloads; runtime versions and TPU generation support must be checked separately.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://docs.cloud.google.com/tpu/docs/pytorch-on-tpu", "checked_at": "2026-11-05", "claim": "Google Cloud documents TorchTPU as an open-source PyTorch-native backend co-developed with the PyTorch team and recommends it for new PyTorch workloads." }, { "type": "upstream_docs", "url": "https://docs.cloud.google.com/tpu/docs/runtimes", "checked_at": "2026-11-05", "claim": "Google Cloud documents generation-specific TPU runtime versions for PyTorch and JAX." } ], "product_ids": [ "tpu7x-ironwood", "tpu-v6e-trillium" ] }, { "id": "google-tpu-vllm", "platform_id": "google-tpu", "layer": "serving", "project": "vLLM-TorchTPU backend", "project_url": "https://docs.cloud.google.com/tpu/docs/tpu-inference", "maintainer_type": "vendor_plugin", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Cloud TPU through the TorchTPU backend; model support and runtime versions are backend- and generation-specific.", "versions": { "serving_engine": "vLLM", "backend": "TorchTPU" }, "limitations": "This is the TPU-specific PyTorch backend, not the CUDA or ROCm path. Confirm the model recipe, TPU generation and runtime version before assuming feature parity.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://docs.cloud.google.com/tpu/docs/tpu-inference", "checked_at": "2026-11-05", "claim": "Google Cloud documents vLLM inference on Cloud TPU through vLLM-TorchTPU for PyTorch models." } ], "product_ids": [ "tpu7x-ironwood", "tpu-v6e-trillium" ] }, { "id": "google-tpu-sglang", "platform_id": "google-tpu", "layer": "serving", "project": "SGLang-TorchTPU backend", "project_url": "https://docs.cloud.google.com/tpu/docs/pytorch-on-tpu", "maintainer_type": "vendor_plugin", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Cloud TPU through the TorchTPU backend; model support and runtime versions are backend- and generation-specific.", "versions": { "serving_engine": "SGLang", "backend": "TorchTPU" }, "limitations": "This is a TPU-specific backend. Confirm model recipes, runtime compatibility and the selected TPU generation; do not infer support from CUDA/ROCm SGLang paths.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://docs.cloud.google.com/tpu/docs/pytorch-on-tpu", "checked_at": "2026-11-05", "claim": "Google Cloud documents SGLang-TorchTPU as a PyTorch backend for Cloud TPU inference." } ], "product_ids": [ "tpu7x-ironwood", "tpu-v6e-trillium" ] }, { "id": "aws-trainium-torch-neuron", "platform_id": "aws-trainium", "layer": "framework", "project": "TorchNeuron", "project_url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/torch/", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Trainium2 (Trn2) and Trainium3 (Trn3) instances.", "versions": { "neuron_sdk": "2.33.0", "release_status": "public beta", "framework": "PyTorch" }, "limitations": "TorchNeuron is documented as public beta in Neuron 2.33.0. Verify operator, distributed feature and model coverage for the exact Neuron release and instance type.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/torch/", "checked_at": "2026-11-05", "claim": "AWS Neuron documentation says TorchNeuron is a public-beta native PyTorch backend supported on Trn2 and Trn3." }, { "type": "upstream_docs", "url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/index.html", "checked_at": "2026-11-05", "claim": "AWS Neuron 2.33.0 documentation describes the Neuron SDK and native PyTorch support for Trainium." } ], "product_ids": [ "trainium2", "trainium3" ] }, { "id": "aws-trainium-vllm-neuron", "platform_id": "aws-trainium", "layer": "serving", "project": "vLLM Neuron plugin", "project_url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/what-is-neuron.html", "maintainer_type": "vendor_plugin", "status": "vendor_plugin_supported", "confidence": "high", "hardware_scope": "Trainium2 (Trn2) and Trainium3 (Trn3) for documented model recipes.", "versions": { "neuron_sdk": "2.33.0", "serving_engine": "vLLM" }, "limitations": "The Neuron plugin provides a Trainium-specific serving path. Model-family recipes and features are version-specific; this record does not imply all upstream vLLM features are supported on all models.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/what-is-neuron.html", "checked_at": "2026-11-05", "claim": "AWS Neuron documents vLLM Neuron as the recommended LLM serving plugin for Trainium2 and Trainium3." }, { "type": "upstream_docs", "url": "https://awsdocs-neuron.readthedocs-hosted.com/en/latest/inference.html", "checked_at": "2026-11-05", "claim": "AWS Neuron documents inference paths and model-specific serving support on Trainium." } ], "product_ids": [ "trainium2", "trainium3" ] }, { "id": "cerebras-pytorch-modelzoo", "platform_id": "cerebras-wse", "layer": "framework", "project": "Cerebras PyTorch and ModelZoo", "project_url": "https://github.com/Cerebras/modelzoo", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Cerebras wafer-scale training systems; exact system generation, deployment mode and model support depend on the documented release.", "versions": { "framework": "PyTorch", "project": "Cerebras ModelZoo" }, "limitations": "Vendor-maintained training framework and model examples. This record documents a public software entry point, not a product-specific benchmark or universal WSE-2/WSE-3 model-compatibility guarantee.", "last_checked": "2026-11-05", "evidence": [ { "type": "repository_readme", "url": "https://github.com/Cerebras/modelzoo", "checked_at": "2026-11-05", "claim": "Cerebras ModelZoo provides PyTorch-based model training workflows for Cerebras systems." }, { "type": "upstream_docs", "url": "https://docs.cerebras.ai/", "checked_at": "2026-11-05", "claim": "Cerebras developer documentation covers training, fine-tuning, PyTorch/ModelZoo and SDK workflows." } ] }, { "id": "cerebras-inference-sdk", "platform_id": "cerebras-wse", "layer": "serving", "project": "Cerebras Inference SDK", "project_url": "https://docs.cerebras.ai/", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Cerebras inference service and supported wafer-scale deployments; exact service SKU and model availability are service-specific.", "versions": { "sdk": "Cerebras Inference SDK" }, "limitations": "Public inference SDK documentation does not imply arbitrary customer control of WSE hardware or identical availability across on-premises and hosted deployments.", "last_checked": "2026-11-05", "evidence": [ { "type": "upstream_docs", "url": "https://docs.cerebras.ai/", "checked_at": "2026-11-05", "claim": "Cerebras developer documentation describes the Inference SDK for building applications on Cerebras inference." } ] }, { "id": "aws-trainium-sglang", "platform_id": "aws-trainium", "layer": "serving", "project": "SGLang AWS Trainium integration", "project_url": "https://github.com/sgl-project/sglang", "maintainer_type": "upstream_project", "status": "in_progress", "confidence": "high", "hardware_scope": "AWS Trainium integration is listed by SGLang as in progress; no Trainium generation or production-ready model coverage is claimed.", "versions": { "release": "v0.5.23 (latest release checked 2026-11-06); platform status read from current upstream README" }, "limitations": "SGLang lists AWS Trainium under integrations in progress, not under supported hardware. Do not treat this as a usable or production-supported path.", "last_checked": "2026-11-06", "evidence": [ { "type": "upstream_docs", "url": "https://github.com/sgl-project/sglang/blob/main/README.md", "checked_at": "2026-11-06", "claim": "SGLang lists AWS Trainium among integrations in progress." } ] }, { "id": "google-tpu-jetstream", "platform_id": "google-tpu", "layer": "serving", "project": "JetStream + Pathways", "project_url": "https://github.com/AI-Hypercomputer/JetStream", "maintainer_type": "vendor", "status": "vendor_framework", "confidence": "high", "hardware_scope": "Google Cloud TPU serving stack. The linked GKE tutorial explicitly demonstrates Llama 3.1-405B on multi-host TPU v6e; do not infer the same recipe on TPU7x without separate evidence.", "versions": { "serving_stack": "JetStream + MaxText + Pathways on GKE", "tutorial_container": "JetStream-Pathways v0.2.0" }, "limitations": "This is a documented Google Cloud recipe, not a universal JetStream model support matrix. The model checkpoint, TPU topology, GKE version and Pathways runtime are part of the configuration.", "last_checked": "2026-11-06", "evidence": [ { "type": "upstream_docs", "url": "https://docs.cloud.google.com/kubernetes-engine/docs/tutorials/serve-multihost-tpu-jetstream", "checked_at": "2026-11-06", "claim": "Google's GKE tutorial deploys Llama 3.1-405B on multi-host TPU v6e using JetStream, MaxText and Pathways." } ], "product_ids": [ "tpu-v6e-trillium" ] } ] }