{ "hf_id": "deepseek-ai/DeepSeek-R1", "meta": { "title": "DeepSeek-R1", "slug": "deepseek-r1", "provider": "DeepSeek", "description": "DeepSeek-R1 is a 671B-parameter MoE reasoning model built on the DeepSeek-V3 architecture, trained with large-scale reinforcement learning for strong chain-of-thought capabilities.", "date_added": "2025-08-01", "date_updated": "2026-10-02", "tasks": [ "text" ], "performance_headline": "Open-weights RL-trained reasoning model with native FP8 / FP4 variants", "related_recipes": [ "deepseek-ai/DeepSeek-V3", "deepseek-ai/DeepSeek-V3.1" ], "hardware": { "h200": "verified", "b200": "verified", "gb200": "verified", "mi300x": "verified", "mi325x": "verified", "mi350x": "verified", "mi355x": "verified" } }, "recommended_command": { "hardware": "h200", "strategy": "single_node_tp", "variant": "default", "node_count": 1, "deploy_type": "single_node", "env": { "VLLM_USE_RUST_FRONTEND": "1" }, "docker_image": "vllm/vllm-openai:latest", "command": "vllm serve deepseek-ai/DeepSeek-R1 \\\n --trust-remote-code \\\n --tensor-parallel-size 8 \\\n --enable-expert-parallel \\\n --enable-auto-tool-choice \\\n --tool-call-parser deepseek_v3 \\\n --chat-template examples/tool_chat_template_deepseekr1.jinja \\\n --reasoning-parser deepseek_r1", "argv": [ "vllm", "serve", "deepseek-ai/DeepSeek-R1", "--trust-remote-code", "--tensor-parallel-size", "8", "--enable-expert-parallel", "--enable-auto-tool-choice", "--tool-call-parser", "deepseek_v3", "--chat-template", "examples/tool_chat_template_deepseekr1.jinja", "--reasoning-parser", "deepseek_r1" ], "docker_command": "docker run --gpus all \\\n --privileged --ipc=host -p 8000:8000 \\\n -v ~/.cache/huggingface:/root/.cache/huggingface \\\n -e VLLM_USE_RUST_FRONTEND=1 \\\n vllm/vllm-openai:latest deepseek-ai/DeepSeek-R1 \\\n --trust-remote-code \\\n --tensor-parallel-size 8 \\\n --enable-expert-parallel \\\n --enable-auto-tool-choice \\\n --tool-call-parser deepseek_v3 \\\n --chat-template examples/tool_chat_template_deepseekr1.jinja \\\n --reasoning-parser deepseek_r1", "docker_argv": [ "docker", "run", "--gpus", "all", "--privileged", "--ipc=host", "-p", "8000:8000", "-v", "~/.cache/huggingface:/root/.cache/huggingface", "-e", "VLLM_USE_RUST_FRONTEND=1", "vllm/vllm-openai:latest", "deepseek-ai/DeepSeek-R1", "--trust-remote-code", "--tensor-parallel-size", "8", "--enable-expert-parallel", "--enable-auto-tool-choice", "--tool-call-parser", "deepseek_v3", "--chat-template", "examples/tool_chat_template_deepseekr1.jinja", "--reasoning-parser", "deepseek_r1" ], "strategy_spec": { "name": "single_node_tp", "deploy_type": "single_node", "display_name": "Tensor Parallel", "short_label": "TP", "orientation": "latency", "description": "Single-node tensor parallel. Splits the model across all local GPUs. TP size is set to the GPU count at deploy time. The simplest multi-GPU strategy — works for all model architectures.\n", "hardware_match": { "min_gpus": 1, "max_gpus": 8, "multi_node": false }, "vllm_args": [], "parallel_flag": "--tensor-parallel-size" }, "hardware_profile": { "brand": "NVIDIA", "generation": "hopper", "compute_arch": "sm90", "display_name": "H200", "description": "NVIDIA H200 SXM 141 GB HBM3e · 8-GPU HGX node", "gpu_count": 8, "vram_gb": 1128, "multi_node": false }, "alternatives": { "single_node_tep": "/deepseek-ai/DeepSeek-R1/strategies/single_node_tep.json", "single_node_dep": "/deepseek-ai/DeepSeek-R1/strategies/single_node_dep.json", "single_node_dpa_tp": "/deepseek-ai/DeepSeek-R1/strategies/single_node_dpa_tp.json", "multi_node_tp": "/deepseek-ai/DeepSeek-R1/strategies/multi_node_tp.json", "multi_node_tp_pp": "/deepseek-ai/DeepSeek-R1/strategies/multi_node_tp_pp.json", "multi_node_dep": "/deepseek-ai/DeepSeek-R1/strategies/multi_node_dep.json", "multi_node_tep": "/deepseek-ai/DeepSeek-R1/strategies/multi_node_tep.json", "pd_cluster": "/deepseek-ai/DeepSeek-R1/strategies/pd_cluster.json", "kv_store_centralized_mooncake": "/deepseek-ai/DeepSeek-R1/strategies/kv_store_centralized_mooncake.json", "kv_store_distributed_mooncake": "/deepseek-ai/DeepSeek-R1/strategies/kv_store_distributed_mooncake.json" }, "by_hardware": { "h200": "/deepseek-ai/DeepSeek-R1/hw/h200.json", "h100": "/deepseek-ai/DeepSeek-R1/hw/h100.json", "b200": "/deepseek-ai/DeepSeek-R1/hw/b200.json", "gb200": "/deepseek-ai/DeepSeek-R1/hw/gb200.json", "b300": "/deepseek-ai/DeepSeek-R1/hw/b300.json", "gb300": "/deepseek-ai/DeepSeek-R1/hw/gb300.json", "mi300x": "/deepseek-ai/DeepSeek-R1/hw/mi300x.json", "mi325x": "/deepseek-ai/DeepSeek-R1/hw/mi325x.json", "mi350x": "/deepseek-ai/DeepSeek-R1/hw/mi350x.json", "mi355x": "/deepseek-ai/DeepSeek-R1/hw/mi355x.json" } }, "model": { "model_id": "deepseek-ai/DeepSeek-R1", "min_vllm_version": "0.12.0", "default_frontend": "rust", "architecture": "moe", "parameter_count": "671B", "active_parameters": "37B", "context_length": 163840, "supports_dcp": true, "base_args": [ "--trust-remote-code" ], "base_env": {}, "install": { "pip": { "command": "uv venv\nsource .venv/bin/activate\nuv pip install -U vllm --torch-backend auto" }, "docker": { "command": "docker pull vllm/vllm-openai:latest" } } }, "features": { "tool_calling": { "description": "Enable DeepSeek-R1 tool calling with the deepseek_v3 tool-call parser.", "args": [ "--enable-auto-tool-choice", "--tool-call-parser", "deepseek_v3", "--chat-template", "examples/tool_chat_template_deepseekr1.jinja" ] }, "reasoning": { "description": "Enable reasoning/thinking mode with the DeepSeek R1 reasoning parser.", "args": [ "--reasoning-parser", "deepseek_r1" ] } }, "opt_in_features": [], "variants": { "default": { "precision": "fp8", "vram_minimum_gb": 805, "description": "Native FP8 weights on 8xH200 (recommended)" }, "r1_0528": { "model_id": "deepseek-ai/DeepSeek-R1-0528", "precision": "fp8", "vram_minimum_gb": 805, "description": "May 2025 DeepSeek-R1 refresh (DeepSeek-R1-0528)", "json": "/deepseek-ai/DeepSeek-R1-0528.json" }, "nvfp4": { "model_id": "nvidia/DeepSeek-R1-0528-NVFP4-v2", "precision": "nvfp4", "vram_minimum_gb": 403, "description": "NVIDIA NVFP4 quantized weights for Blackwell GPUs", "extra_args": [ "--kv-cache-dtype", "fp8" ], "json": "/nvidia/DeepSeek-R1-0528-NVFP4-v2.json" }, "mxfp4": { "model_id": "amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4", "precision": "mxfp4", "vram_minimum_gb": 403, "description": "AMD MXFP4 quantized weights for MI350X / MI355X (gfx950)", "supported_hardware": [ "mi350x", "mi355x" ], "tp": 8, "extra_args": [ "--no-enable-expert-parallel", "--kv-cache-dtype", "fp8" ], "extra_env": { "VLLM_ROCM_USE_SKINNY_GEMM": "0" }, "json": "/amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4.json" }, "redhat_nvfp4": { "model_id": "RedHatAI/DeepSeek-R1-NVFP4-FP8-BLOCK", "precision": "nvfp4", "label": "NVFP4+FP8 Block (RedHat)", "vram_minimum_gb": 444, "description": "RedHatAI NVFP4+FP8 block mixed-precision checkpoint", "json": "/RedHatAI/DeepSeek-R1-NVFP4-FP8-BLOCK.json" } }, "hardware_overrides": { "nvidia": { "extra_args": [ "--enable-expert-parallel" ], "extra_env": {} }, "amd": { "extra_args": [], "extra_env": { "VLLM_ROCM_USE_AITER": "1" } } }, "guide": "## Overview\n\nDeepSeek-R1 is a 671B-parameter Mixture-of-Experts reasoning model (37B activated per\ntoken) that shares its architecture with DeepSeek-V3, so the same launch recipes apply\nto both. DeepSeek publishes a refreshed checkpoint as `DeepSeek-R1-0528`, NVIDIA\npublishes an FP4 quantized variant (`nvidia/DeepSeek-R1-FP4`) that runs on Blackwell\nGPUs with fewer devices, and AMD publishes an MXFP4 quantized variant\n(`amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4`) that runs on Instinct GPUs.\n\n## Prerequisites\n\n- **Hardware (FP8)**: 8x H200 (CUDA) or 8x MI300X / MI325X / MI355X (ROCm)\n- **Hardware (NVFP4)**: 4x B200 GPUs\n- **Hardware (MXFP4)**: 8x MI350X / MI355X (ROCm)\n- **vLLM**: Current stable release\n\n### CUDA\n```bash\nuv venv\nsource .venv/bin/activate\nuv pip install -U vllm --torch-backend auto\n```\n\n### ROCm (MI300X, MI325X, MI355X)\nRequires Python 3.12, ROCm 7.2.1, and glibc >= 2.35.\n```bash\nuv venv --python 3.12\nsource .venv/bin/activate\nuv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/\n```\n\n## Client Usage\n\n### 8xH200 / 8xMI300X (FP8)\n\nTensor Parallel + Expert Parallel (TP8+EP) — CUDA:\n```bash\nvllm serve deepseek-ai/DeepSeek-R1-0528 \\\n --trust-remote-code \\\n --tensor-parallel-size 8 \\\n --enable-expert-parallel\n```\n\nTensor Parallel + Expert Parallel (TP8+EP) — ROCm:\n```bash\nexport VLLM_ROCM_USE_AITER=1\n\nvllm serve deepseek-ai/DeepSeek-R1 \\\n --trust-remote-code \\\n --tensor-parallel-size 8 \\\n --enable-expert-parallel\n```\n\nData Parallel + Expert Parallel (DP8+EP) — CUDA:\n```bash\nvllm serve deepseek-ai/DeepSeek-R1-0528 \\\n --trust-remote-code \\\n --data-parallel-size 8 \\\n --enable-expert-parallel\n```\n\n### 8xMI355X (FP8)\n\nTensor Parallel (TP8) — ROCm. Validated on MI355X only; INT4 quick reduce and\nfused shared experts are not enabled on other AMD GPUs or other strategies:\n```bash\nexport VLLM_ROCM_USE_AITER=1\nexport VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4\nexport VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1\n\nvllm serve deepseek-ai/DeepSeek-R1 \\\n --trust-remote-code \\\n --tensor-parallel-size 8\n```\n\n### 8xMI350X / 8xMI355X (DPA+TP)\n\nEach GPU runs its own attention replica with a private KV cache, while the MoE\nexperts stay tensor-sharded across all eight GPUs. Aggregate KV-cache\ncapacity is roughly 8x that of a single TP8 replica, useful for workloads \nbenefiting from larger KV capacities. \n\nEngine (all eight ranks in one process, listening on 8100):\n```bash\nexport VLLM_ROCM_USE_AITER=1\nexport VLLM_ENGINE_READY_TIMEOUT_S=1800\n\nvllm serve deepseek-ai/DeepSeek-R1 \\\n --trust-remote-code \\\n --port 8100 \\\n --data-parallel-size 8 \\\n --tensor-parallel-size 1\n```\n\nRouter (client-facing on 8000, spreads requests across the eight DP ranks):\n```bash\nuv pip install vllm-router\n\nvllm-router \\\n --host 0.0.0.0 \\\n --port 8000 \\\n --policy consistent_hash \\\n --intra-node-data-parallel-size 8 \\\n --worker-urls http://localhost:8100 \\\n --worker-startup-timeout-secs 1800\n```\n\nNotes:\n\n- `--intra-node-data-parallel-size` must match the engine's\n `--data-parallel-size`. A mismatch silently under-uses ranks.\n- `consistent_hash` is the policy that makes this worthwhile for chat: it pins\n a conversation to the rank that already holds its prefix, so the larger\n aggregate cache actually gets hit. `round_robin` scatters turns of the same\n conversation across ranks and throws the prefix away.\n- Start the engine first, then the router; `--worker-startup-timeout-secs 1800`\n covers the engine's weight load either way.\n- Send client traffic to the router on port 8000, never to 8100 directly.\n\n### 4xB200 (FP4)\n\nOn vLLM v0.28.0 and later, the FlashInfer MoE kernels are selected automatically\nfor both FP4 and FP8 on Blackwell — no environment variables are needed. To pin\nthe backend explicitly, pass `--moe-backend flashinfer_trtllm`.\n\nTensor Parallel + Expert Parallel (TP4+EP):\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 vllm serve nvidia/DeepSeek-R1-FP4 \\\n --trust-remote-code \\\n --tensor-parallel-size 4 \\\n --enable-expert-parallel\n```\n\nData Parallel + Expert Parallel (DP4+EP):\n```bash\nCUDA_VISIBLE_DEVICES=0,1,2,3 vllm serve nvidia/DeepSeek-R1-FP4 \\\n --trust-remote-code \\\n --data-parallel-size 4 \\\n --enable-expert-parallel\n```\n\n### 8xMI350X / 8xMI355X (MXFP4)\n\nTensor Parallel + FP8 KV cache (TP8):\n```bash\nexport VLLM_ROCM_USE_AITER=1\nexport VLLM_ROCM_USE_SKINNY_GEMM=0\n\nvllm serve amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4 \\\n --trust-remote-code \\\n --tensor-parallel-size 8 \\\n --kv-cache-dtype fp8\n```\n\n## Benchmarking\n\nFor benchmarking, prefix caching is disabled by default in vLLM — no extra server\nflag is needed.\n\n```bash\n# FP8 benchmark\nvllm bench serve \\\n --model deepseek-ai/DeepSeek-R1-0528 \\\n --dataset-name random \\\n --random-input-len 8000 \\\n --random-output-len 1000 \\\n --request-rate 10000 \\\n --num-prompts 16 \\\n --ignore-eos\n```\n\n```bash\n# NVFP4 benchmark\nvllm bench serve \\\n --model nvidia/DeepSeek-R1-FP4 \\\n --dataset-name random \\\n --random-input-len 8000 \\\n --random-output-len 1000 \\\n --request-rate 10000 \\\n --num-prompts 16 \\\n --ignore-eos\n```\n\n```bash\n# MXFP4 benchmark (MI350X / MI355X)\nvllm bench serve \\\n --model amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4 \\\n --dataset-name random \\\n --random-input-len 8000 \\\n --random-output-len 1000 \\\n --request-rate 10000 \\\n --num-prompts 16 \\\n --ignore-eos\n```\n\nTest different workloads by adjusting input/output lengths:\n- Prompt-heavy: 8000 input / 1000 output\n- Decode-heavy: 1000 input / 8000 output\n- Balanced: 1000 input / 1000 output\n\n## Troubleshooting\n\n- **Disaggregated Serving with Wide EP (Experimental GB200)**: See\n [vLLM issue #33583](https://github.com/vllm-project/vllm/issues/33583),\n [the vLLM blog post](https://blog.vllm.ai/2026/02/03/dsr1-gb200-part1.html), and\n [this reference fork](https://github.com/minosfuture/vllm/tree/pd_gb200_0114/runs/DS-R1/fp4)\n for GB200 disaggregated serving recipes.\n\n## References\n\n- [DeepSeek-R1 on Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-R1)\n- [DeepSeek-R1-0528 on Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528)\n- [NVIDIA DeepSeek-R1-FP4](https://huggingface.co/nvidia/DeepSeek-R1-FP4)\n- [AMD DeepSeek-R1-0528-MXFP4-MTP-MoEFP4](https://huggingface.co/amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4)\n- [vLLM Expert Parallelism docs](https://docs.vllm.ai/en/latest/serving/expert_parallel_deployment.html)\n- [vLLM Data Parallel deployment docs](https://docs.vllm.ai/en/latest/serving/data_parallel_deployment.html)\n", "source": { "source": "vllm", "repo": "vllm-project/vllm", "ref": "main", "sha": "564ac0de1b228568cafcbfbe8ccaccc82d6023eb", "path": "recipes/models/deepseek-ai/DeepSeek-R1.yaml", "edit_url": "https://github.com/vllm-project/vllm/edit/main/recipes/models/deepseek-ai/DeepSeek-R1.yaml", "overlays": [] } }