{
  "schema_version": "1.0",
  "database_version": "2026.11.05-2",
  "last_reviewed": "2026-11-05",
  "scope": "Workload taxonomy and candidate hardware/software execution paths. A path may only document a software integration or vendor report; it is not a certification of workload support. Only explicit workload tests or benchmark records can establish workload-specific validation.",
  "status_definitions": {
    "integration_only": "A software-to-platform integration is documented, but no workload-specific model/configuration validation is attached.",
    "vendor_reported": "A vendor publicly reports a workload or model adaptation; independent reproduction and benchmark comparability are not established.",
    "workload_tested": "A specific workload configuration has passed a documented correctness or functional test; performance is not implied.",
    "benchmark_reported": "A workload benchmark result is published with configuration and measurement metadata, but is not independently reproduced here.",
    "independently_benchmarked": "A benchmark result has been independently reproduced under a documented methodology."
  },
  "workloads": [
    {
      "id": "llm-serving",
      "name": "LLM inference & serving",
      "category": "llm_inference",
      "phase": "online_serving",
      "description": "End-to-end online generation under a request-serving policy. Keep prefill and decode behavior visible rather than reducing the workload to a single tokens-per-second number.",
      "representative_models": [
        "Decoder-only transformer LLMs",
        "Dense and mixture-of-experts models (pin the exact architecture and revision in each test)"
      ],
      "key_dimensions": [
        "Prompt and output token lengths",
        "Concurrency and arrival pattern",
        "Batching and scheduling policy",
        "KV-cache capacity, reuse and offload",
        "Quantization and attention backend",
        "TP / PP / EP topology and interconnect"
      ],
      "metrics_to_report": [
        "TTFT",
        "TPOT / inter-token latency",
        "Request throughput and output tokens/s",
        "P50 / P95 / P99 latency",
        "Goodput under an SLO",
        "Cost per million output tokens"
      ]
    },
    {
      "id": "llm-prefill",
      "name": "LLM prefill",
      "category": "llm_inference",
      "phase": "prefill",
      "description": "Prompt ingestion and context processing. Compare compute throughput, memory traffic and time-to-first-token under explicitly stated prompt lengths.",
      "representative_models": [
        "Decoder-only transformer prefill"
      ],
      "key_dimensions": [
        "Input token length",
        "Chunked prefill policy",
        "Batch size and concurrency",
        "Attention implementation",
        "Prefix / prompt cache hit rate",
        "Weight and activation precision"
      ],
      "metrics_to_report": [
        "Prefill tokens/s",
        "Time to first token",
        "Latency by input length",
        "Peak memory and bandwidth utilization"
      ]
    },
    {
      "id": "llm-decode",
      "name": "LLM decode",
      "category": "llm_inference",
      "phase": "decode",
      "description": "Autoregressive token generation, where KV-cache movement, memory bandwidth, batching and communication often dominate.",
      "representative_models": [
        "Dense and MoE autoregressive LLM decode"
      ],
      "key_dimensions": [
        "KV-cache sequence length",
        "Active sequences and decode batch",
        "KV-cache dtype and layout",
        "Speculative decoding method",
        "TP / EP and all-to-all topology",
        "Memory bandwidth and cache residency"
      ],
      "metrics_to_report": [
        "TPOT / TBT",
        "Output tokens/s per device and per system",
        "Concurrent sequences at target latency",
        "KV-cache bytes per token",
        "Interconnect traffic and utilization"
      ]
    },
    {
      "id": "llm-pretraining",
      "name": "LLM pretraining",
      "category": "llm_training",
      "phase": "pretraining",
      "description": "Large-scale distributed training from initialization. A valid comparison must preserve model shape, token budget, parallelism and convergence target.",
      "representative_models": [
        "Dense transformer pretraining",
        "Mixture-of-experts pretraining"
      ],
      "key_dimensions": [
        "Parameter count and active parameters",
        "Training tokens and sequence length",
        "DP / TP / PP / CP / EP topology",
        "Optimizer and precision",
        "Checkpointing and recomputation",
        "Network topology and collective implementation"
      ],
      "metrics_to_report": [
        "Tokens/s and time-to-train",
        "MFU / HFU with stated definition",
        "Step time and scaling efficiency",
        "Communication / compute overlap",
        "Energy and cost per training token"
      ]
    },
    {
      "id": "llm-finetuning",
      "name": "LLM fine-tuning",
      "category": "llm_training",
      "phase": "finetuning",
      "description": "Full-parameter and parameter-efficient fine-tuning. Separate software-framework availability from validated recipes, convergence and end-to-end throughput.",
      "representative_models": [
        "Supervised fine-tuning",
        "LoRA / QLoRA",
        "Preference optimization and post-training"
      ],
      "key_dimensions": [
        "Trainable parameter fraction",
        "Sequence length and batch size",
        "Optimizer state and activation memory",
        "FSDP / ZeRO / TP strategy",
        "Precision and quantization",
        "Dataset and convergence criteria"
      ],
      "metrics_to_report": [
        "Training tokens/s",
        "Step time and time-to-quality",
        "Peak memory",
        "Scaling efficiency",
        "Cost per fine-tuning run"
      ]
    },
    {
      "id": "llm-model-porting",
      "name": "LLM model porting & adaptation",
      "category": "model_porting",
      "phase": "model_porting",
      "description": "Port a model architecture and its operators to a new hardware/software stack. A model loading demo is not equivalent to a validated serving or benchmark result.",
      "representative_models": [
        "New model architecture or model revision",
        "Vendor-specific accelerator adaptation"
      ],
      "key_dimensions": [
        "Operator coverage and fallback ratio",
        "Numerical parity and accuracy tolerance",
        "Attention / MoE / quantization kernels",
        "Checkpoint conversion and weight layout",
        "Runtime and compiler versions",
        "Reproducible launch and test recipe"
      ],
      "metrics_to_report": [
        "Operator pass rate",
        "Numerical parity / task quality",
        "Model load and first-token success",
        "Fallback ratio",
        "End-to-end throughput after correctness passes"
      ]
    },
    {
      "id": "multimodal-serving",
      "name": "Multimodal / VLM serving",
      "category": "multimodal_inference",
      "phase": "multimodal_serving",
      "description": "Online inference for models combining text with image, audio or video inputs. Preserve modality-specific preprocessing and sequence expansion.",
      "representative_models": [
        "Vision-language models",
        "Audio-language models",
        "Video-language models"
      ],
      "key_dimensions": [
        "Image resolution / patch count",
        "Audio duration and sampling rate",
        "Video frame count and sampling policy",
        "Text context and generated tokens",
        "Preprocessing placement",
        "Multimodal KV-cache and batching"
      ],
      "metrics_to_report": [
        "End-to-end latency by modality",
        "Input items/s and output tokens/s",
        "TTFT / TPOT",
        "Preprocessing and encoder latency",
        "Quality under a fixed evaluation set"
      ]
    },
    {
      "id": "embedding-reranking",
      "name": "Embedding & reranking",
      "category": "retrieval",
      "phase": "retrieval",
      "description": "Vector embedding and relevance reranking for retrieval-augmented generation and search. Compare throughput only at a fixed model, sequence length and quality target.",
      "representative_models": [
        "Text embedding models",
        "Cross-encoder rerankers",
        "Multilingual retrieval models"
      ],
      "key_dimensions": [
        "Query / document length",
        "Vector dimension",
        "Batch size and concurrency",
        "Pooling and normalization",
        "Index and retrieval implementation",
        "Quality dataset and recall target"
      ],
      "metrics_to_report": [
        "Queries/s and documents/s",
        "P50 / P95 latency",
        "Recall@K / NDCG",
        "Throughput at a fixed quality target",
        "Memory footprint"
      ]
    },
    {
      "id": "diffusion-generation",
      "name": "Diffusion image & video generation",
      "category": "generative_inference",
      "phase": "generative_inference",
      "description": "Image and video generation pipelines. Separate text encoder, denoising, VAE and video temporal components when reporting latency.",
      "representative_models": [
        "Text-to-image diffusion",
        "Image editing and image-to-video",
        "Video diffusion models"
      ],
      "key_dimensions": [
        "Resolution and frame count",
        "Denoising steps and sampler",
        "Batch size",
        "Precision and quantization",
        "Attention / convolution kernels",
        "Pipeline and model parallelism"
      ],
      "metrics_to_report": [
        "Seconds per image / video",
        "Images/s or frames/s",
        "Peak memory",
        "Quality score under a fixed evaluator",
        "Energy per generated output"
      ]
    },
    {
      "id": "edge-vision",
      "name": "Edge vision & perception",
      "category": "edge_inference",
      "phase": "edge_inference",
      "description": "On-device and embedded vision workloads under power, thermal, memory and real-time constraints.",
      "representative_models": [
        "Image classification",
        "Object detection and segmentation",
        "Video perception"
      ],
      "key_dimensions": [
        "Input resolution and frame rate",
        "Batch size and real-time deadline",
        "Precision and sparsity",
        "Host-device data movement",
        "Thermal envelope and cooling",
        "Camera / sensor preprocessing"
      ],
      "metrics_to_report": [
        "FPS and end-to-end latency",
        "Power and energy per frame",
        "Accuracy / mAP",
        "Memory footprint",
        "Sustained performance under thermal limits"
      ]
    }
  ],
  "paths": [
    {
      "id": "path-nvidia-vllm",
      "platform_id": "nvidia-gpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "nvidia-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-nvidia-sglang",
      "platform_id": "nvidia-gpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "nvidia-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-amd-vllm",
      "platform_id": "amd-gpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "amd-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-amd-sglang",
      "platform_id": "amd-gpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "amd-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-intel-vllm",
      "platform_id": "intel-xpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "intel-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-intel-sglang",
      "platform_id": "intel-xpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "intel-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-ascend-vllm",
      "platform_id": "huawei-ascend",
      "workload_id": "llm-serving",
      "integration_ids": [
        "ascend-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-ascend-sglang",
      "platform_id": "huawei-ascend",
      "workload_id": "llm-serving",
      "integration_ids": [
        "ascend-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-mthreads-vllm",
      "platform_id": "moore-threads",
      "workload_id": "llm-serving",
      "integration_ids": [
        "mthreads-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-mthreads-sglang",
      "platform_id": "moore-threads",
      "workload_id": "llm-serving",
      "integration_ids": [
        "mthreads-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-metax-vllm",
      "platform_id": "metax",
      "workload_id": "llm-serving",
      "integration_ids": [
        "metax-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-metax-sglang",
      "platform_id": "metax",
      "workload_id": "llm-serving",
      "integration_ids": [
        "metax-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-hygon-sglang",
      "platform_id": "hygon",
      "workload_id": "llm-serving",
      "integration_ids": [
        "hygon-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-cambricon-sglang",
      "platform_id": "cambricon",
      "workload_id": "llm-serving",
      "integration_ids": [
        "cambricon-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-iluvatar-vllm",
      "platform_id": "iluvatar",
      "workload_id": "llm-serving",
      "integration_ids": [
        "iluvatar-vllm"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-iluvatar-sglang",
      "platform_id": "iluvatar",
      "workload_id": "llm-serving",
      "integration_ids": [
        "iluvatar-sglang"
      ],
      "status": "integration_only",
      "scope": "The cited source documents a software integration for this accelerator. It does not establish end-to-end support for a specific model, precision, context length, parallelism mode or workload feature.",
      "limitations": "No workload-specific model recipe, device count, precision, sequence-length configuration, correctness run or performance measurement is attached to this path yet.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-biren-deepseek-v4-port",
      "platform_id": "biren",
      "workload_id": "llm-model-porting",
      "integration_ids": [
        "biren-deepseek-v4-port"
      ],
      "status": "vendor_reported",
      "scope": "Vendor-reported adaptation of a DeepSeek-V4 preview model on BR166. The public claim is a model-porting report, not an independently reproducible benchmark or proof of a general-purpose serving backend.",
      "limitations": "No reproducible batch / sequence / precision configuration, correctness protocol, performance metrics or independent verification is attached to the current public record.",
      "configuration": {
        "model": "DeepSeek-V4 preview",
        "device_scope": "Biren BR166"
      },
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-intel-gaudi-vllm",
      "platform_id": "intel-gaudi",
      "workload_id": "llm-serving",
      "integration_ids": [
        "intel-gaudi-vllm"
      ],
      "status": "integration_only",
      "scope": "Gaudi documentation lists a vLLM hardware-plugin path for Gaudi 2 and Gaudi 3.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-google-tpu-torchpytorch-pretraining",
      "platform_id": "google-tpu",
      "workload_id": "llm-pretraining",
      "integration_ids": [
        "google-tpu-torchtpu"
      ],
      "status": "integration_only",
      "scope": "Google Cloud documents PyTorch training on Cloud TPU through TorchTPU; this is an integration record, not a measured training result.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-google-tpu-torchpytorch-finetuning",
      "platform_id": "google-tpu",
      "workload_id": "llm-finetuning",
      "integration_ids": [
        "google-tpu-torchtpu"
      ],
      "status": "integration_only",
      "scope": "Google Cloud documents fine-tuning open models on Cloud TPU through its PyTorch-native TPU stack; no product-specific benchmark is attached.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-google-tpu-vllm-serving",
      "platform_id": "google-tpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "google-tpu-vllm"
      ],
      "status": "integration_only",
      "scope": "Google Cloud documents vLLM serving on Cloud TPU through the TorchTPU backend; exact model recipes and TPU generation support vary.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-google-tpu-sglang-serving",
      "platform_id": "google-tpu",
      "workload_id": "llm-serving",
      "integration_ids": [
        "google-tpu-sglang"
      ],
      "status": "integration_only",
      "scope": "Google Cloud documents SGLang serving on Cloud TPU through the TorchTPU backend; exact model recipes and TPU generation support vary.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-aws-trainium-torchneuron-pretraining",
      "platform_id": "aws-trainium",
      "workload_id": "llm-pretraining",
      "integration_ids": [
        "aws-trainium-torch-neuron"
      ],
      "status": "integration_only",
      "scope": "AWS Neuron documents PyTorch training on Trainium through TorchNeuron; TorchNeuron is public beta in Neuron 2.33.0.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-aws-trainium-torchneuron-finetuning",
      "platform_id": "aws-trainium",
      "workload_id": "llm-finetuning",
      "integration_ids": [
        "aws-trainium-torch-neuron"
      ],
      "status": "integration_only",
      "scope": "AWS Neuron documents training and fine-tuning paths on Trainium; this record does not assert a workload test or benchmark.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-aws-trainium-vllm-serving",
      "platform_id": "aws-trainium",
      "workload_id": "llm-serving",
      "integration_ids": [
        "aws-trainium-vllm-neuron"
      ],
      "status": "integration_only",
      "scope": "AWS Neuron documents vLLM Neuron as an LLM serving path for Trainium2 and Trainium3; model coverage is recipe-specific.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-cerebras-pytorch-pretraining",
      "platform_id": "cerebras-wse",
      "workload_id": "llm-pretraining",
      "integration_ids": [
        "cerebras-pytorch-modelzoo"
      ],
      "status": "integration_only",
      "scope": "Cerebras documents PyTorch and ModelZoo training workflows on wafer-scale systems; no measured result is attached to this path.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-cerebras-pytorch-finetuning",
      "platform_id": "cerebras-wse",
      "workload_id": "llm-finetuning",
      "integration_ids": [
        "cerebras-pytorch-modelzoo"
      ],
      "status": "integration_only",
      "scope": "Cerebras documents fine-tuning workflows through its PyTorch and ModelZoo stack; hardware and model coverage are release-specific.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    },
    {
      "id": "path-cerebras-inference",
      "platform_id": "cerebras-wse",
      "workload_id": "llm-serving",
      "integration_ids": [
        "cerebras-inference-sdk"
      ],
      "status": "integration_only",
      "scope": "Cerebras documents an Inference SDK; this path records public software entry points, not a product-specific benchmark.",
      "limitations": "This is a documented software/workload path only. No product-specific correctness run, configuration, throughput, latency, power or independent benchmark result is attached. Do not interpret this as workload-tested or benchmarked support.",
      "configuration": {},
      "metrics": [],
      "last_checked": "2026-11-05"
    }
  ]
}