{ "schema_version": 1, "deployment_id": "glm53-uncensored-nvfp4-tp16-high", "title": "GLM-5.3 modified NVFP4, sixteen Sparks", "model": { "repository": "dealignai/GLM-5.3-UNCENSORED-NVFP4", "revision": "409ddf484d16cb7b13d390e8b30942890164d5af", "url": "https://huggingface.co/dealignai/GLM-5.3-UNCENSORED-NVFP4", "revision_url": "https://huggingface.co/dealignai/GLM-5.3-UNCENSORED-NVFP4/tree/409ddf484d16cb7b13d390e8b30942890164d5af", "license_declaration": "Modified card declares MIT; its base card declares custom GLM5.3 license", "verification": "Complete Sep19 checkpoint SHA256 receipts; metadata and mount identity checked", "download_argv": [ "hf", "download", "dealignai/GLM-5.3-UNCENSORED-NVFP4", "--revision", "409ddf484d16cb7b13d390e8b30942890164d5af", "--local-dir", "{MODEL_DIR}" ] }, "hardware": { "node_count": 16, "gpu_count": 16, "accelerator": "NVIDIA GB10 DGX Spark", "memory_per_gpu_gib": 128, "memory_type": "unified", "interconnect": "Dual ConnectX-7 RoCEv2 through switched fabric", "note": "This describes tested hardware, not a minimum requirement or a throughput guarantee." }, "public_download_path": "/benchmarks/recipes/glm53-uncensored-nvfp4-tp16-high.json", "record_scope": "Observed serving configuration; independent reproduction not yet performed", "reproduction": { "status": "public-source-build-recipe-available", "limitations": [ "A clean Dockerfile and exact upstream method repair are supplied. The script reproduces the verified source-file hash; a clean image build and independent inference reproduction have not been performed.", "The campaign image was assembled as a deterministic delta. A conventional Docker build has different metadata and image ID; retain original campaign identities.", "Argument snapshot includes the checkpoint architecture index pattern; retain the exact pattern." ] }, "sources": [ "https://github.com/ciprianveg/gb10-glm-5.2", "https://github.com/im0xMagnus/glm-5.3-uncensored-8x-dgx-spark/tree/dd5fff7a1472d8c38d0b624422d5c7e7db20138d", "https://github.com/vllm-project/vllm/pull/44993", "https://huggingface.co/dealignai/GLM-5.3-UNCENSORED-NVFP4/tree/409ddf484d16cb7b13d390e8b30942890164d5af" ], "placeholders": { "MODEL_DIR": "Checkpoint snapshot directory on each rank or a read-only shared filesystem", "DRAFT_DIR": "Separate pinned draft checkpoint directory when required", "CACHE_DIR": "Writable runtime cache", "NODE_RANK": "Zero-based rank", "NODE_IP": "This node fabric address", "HEAD_IP": "Rank-zero fabric address", "BIND_IP": "Desired API bind address", "API_PORT": "API listen port", "DIST_PORT": "Unused distributed coordination port", "FABRIC_INTERFACE": "Your fabric interface", "RDMA_DEVICES": "Your active RoCE device list", "SERVED_MODEL_NAME": "Public API model name chosen by the deployer" }, "limitations": [ "Configured context and scheduler slots are capacity limits, not proof of simultaneous useful context or latency.", "Use identical pinned weights/runtime across ranks and validate driver, CUDA architecture and RDMA interface compatibility.", "Effort labels have model-specific meanings; high and medium are not cross-model compute units.", "Server default effort was low; all benchmark requests explicitly overrode it to high." ], "runtime": { "engine": "vLLM with merged upstream structured-output repair", "version": "v19 + upstream PR44993", "engine_source_revision": "264bce1da81e27d638e7cf265b4cbd125d023c38 (base image label; layered overlays also apply)", "image_id": "sha256:a42466afd9252a63c17dacc86578c9b6046c1f0456920c78aaa618c66fd23bf0", "pull_reference": null, "base_pull_reference": "ghcr.io/ciprianveg/gb10-glm-5.2@sha256:7ac0031cfefc83e2d3fb127faf0c4a58ba25c8a33e2e5690c3e3760d51097898", "source_url": "https://github.com/ciprianveg/gb10-glm-5.2", "build_instructions_url": "/benchmarks/recipes/builds/glm53-grammar/README.md", "upstream_recipe_url": "https://github.com/im0xMagnus/glm-5.3-uncensored-8x-dgx-spark/tree/dd5fff7a1472d8c38d0b624422d5c7e7db20138d", "runtime_patch": { "upstream_url": "https://github.com/vllm-project/vllm/pull/44993", "merge_commit": "0416dab275d51327b331a1c6baaec754a68d7764", "before_sha256": "4bddb36d3e424946b095b0fc4a6d91137a4ead0e08c00f00a40d9258d15d70c1", "after_sha256": "0f72f8f819262b298d95b7333deb7ec0a89aa9c57cce81b6c793429eb915b33d", "operation": "Replace should_advance with exact merged upstream method; scheduler trim hook and boundary helper already match." }, "build_bundle_manifest_url": "/benchmarks/recipes/builds/manifest.json", "dockerfile_url": "/benchmarks/recipes/builds/glm53-grammar/Dockerfile", "build_bundle_url": "/benchmarks/recipes/builds/glm53-grammar.tar.gz", "build_bundle_sha256": "a052de8e603c46a2dccff9c255b3a81cfb623e447c558561e0516aa80e99fc02" }, "serving": { "executable": "vllm", "argv": [ "serve", "/model", "--served-model-name", "{SERVED_MODEL_NAME}", "--host", "{BIND_IP}", "--port", "{API_PORT}", "--tensor-parallel-size", "16", "--pipeline-parallel-size", "1", "--decode-context-parallel-size", "1", "--nnodes", "16", "--node-rank", "{NODE_RANK}", "--master-addr", "{HEAD_IP}", "--master-port", "{DIST_PORT}", "--max-model-len", "1048576", "--max-num-seqs", "16", "--max-num-batched-tokens", "4096", "--gpu-memory-utilization", "0.78", "--kv-cache-dtype", "fp8_ds_mla", "--speculative-config", "{\"model\": \"/model\", \"draft_attention_backend\": \"B12X_MLA_SPARSE\", \"method\": \"mtp\", \"num_speculative_tokens\": 3, \"draft_tensor_parallel_size\": 1}", "--default-chat-template-kwargs", "{\"reasoning_effort\": \"low\"}", "--attention-backend", "B12X_MLA_SPARSE", "--hf-overrides", "{\"index_topk_pattern\": \"FFFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSSFSSS\"}", "--dcp-kv-cache-interleave-size", "1", "--enable-chunked-prefill", "--enable-prefix-caching", "--async-scheduling", "--long-prefill-token-threshold", "2048", "--enable-decode-aware-prefill", "--decode-prefill-token-budget", "1024", "--idle-prefill-token-budget", "16384", "--max-long-prefills-per-step", "1", "--compilation-config", "{\"cudagraph_mode\": \"FULL\", \"max_cudagraph_capture_size\": 64}", "--disable-custom-all-reduce", "--reasoning-parser", "glm45", "--tool-call-parser", "glm47", "--enable-auto-tool-choice", "--generation-config", "vllm", "--distributed-executor-backend", "mp", "--seed", "42" ], "environment": { "VLLM_ENGINE_READY_TIMEOUT_S": "7200", "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "1800", "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True", "TORCH_CUDA_ARCH_LIST": "12.1a", "CUTE_DSL_ARCH": "sm_121a", "VLLM_USE_V2_MODEL_RUNNER": "1", "VLLM_USE_B12X_SPARSE_INDEXER": "1", "VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": "0", "DETERMINISTIC_MOE_ALIGN": "1", "VLLM_MARLIN_USE_ATOMIC_ADD": "1", "VLLM_DCP_QUERY_SPLIT": "1", "VLLM_B12X_MLA_CKG_GATHER": "1", "OMP_NUM_THREADS": "8", "VLLM_MTP_INSTRUMENT": "1", "VLLM_MTP_INSTRUMENT_WINDOW": "32", "VLLM_B12X_MLA_SPEC_EXTEND_AS_DECODE": "1", "VLLM_SPARSE_INDEXER_MAX_LOGITS_MB": "256", "GLM52_BIND_HOST_TRITON": "0", "HF_HOME": "/cache/huggingface", "HF_HUB_OFFLINE": "1", "TRANSFORMERS_OFFLINE": "1", "TORCH_EXTENSIONS_DIR": "/cache/torch_extensions", "TORCHINDUCTOR_CACHE_DIR": "/cache/torchinductor-cache", "TRITON_CACHE_DIR": "/cache/triton-cache", "XDG_CACHE_HOME": "/cache", "B12X_CUTE_COMPILE_CACHE_DIR": "/cache/b12x/cute_compile", "B12X_LOG_CUTE_COMPILES_AFTER_ENGINE_START": "0", "NCCL_NET": "IB", "NCCL_IB_DISABLE": "0", "NCCL_IB_ROCE_VERSION_NUM": "2", "NCCL_IB_ADDR_FAMILY": "AF_INET", "NCCL_CROSS_NIC": "1", "NCCL_NVLS_ENABLE": "0", "NCCL_CUMEM_ENABLE": "0", "NCCL_IGNORE_CPU_AFFINITY": "1", "NCCL_BUFFSIZE": "16777216", "NCCL_MAX_NCHANNELS": "4", "NCCL_MIN_NCHANNELS": "4", "NCCL_DEBUG": "WARN", "TORCH_NCCL_ASYNC_ERROR_HANDLING": "1", "VLLM_HOST_IP": "{NODE_IP}", "NCCL_SOCKET_IFNAME": "{FABRIC_INTERFACE}", "GLOO_SOCKET_IFNAME": "{FABRIC_INTERFACE}", "TP_SOCKET_IFNAME": "{FABRIC_INTERFACE}", "MN_IF_NAME": "{FABRIC_INTERFACE}", "NCCL_IB_HCA": "={RDMA_DEVICES}", "NCCL_IB_GID_INDEX": "3" }, "tp": 16, "ep": "Not enabled", "dcp": 1, "context_tokens": 1048576, "scheduler_slots": 16, "weight_format": "NVFP4", "kv_format": "FP8 MLA", "speculation": { "method": "MTP", "draft_tokens": 3, "draft_tp": 1, "draft_checkpoint": "Same checkpoint" }, "reasoning_parser": "glm45", "tool_parser": "glm47", "request_body": { "chat_template_kwargs": { "reasoning_effort": "high" } }, "container_mounts": [ "{MODEL_DIR}:/model:ro", "{CACHE_DIR}:/cache" ], "worker_rank_extra_argv": [ "--headless" ], "container_entrypoint_override": "vllm" } }