{ "schema": "anvil-serving.deepseek-r33-qualification-status/v1", "updated_at_utc": "2026-08-11T04:30:00Z", "checkpoint": { "repository": "deepseek-ai/DeepSeek-V4-Flash-0731", "revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb", "substitution_allowed": false }, "hardware": { "topology": "two equal RTX PRO 6000 Blackwell Max-Q GPUs, exclusive TP=2 over PCIe without NVLink", "aggregate_vram_gib": 192, "aggregate_vram_is_unified": false }, "operator_priority": [ "quality", "stability", "performance" ], "memory_policy": { "mode": "exclusive-ai-only", "video_workload_vram_reserve_mib_per_gpu": 0, "hard_vram_reserve_gate": false, "vllm_gpu_memory_utilization": 0.975, "allocator_note": "The vLLM fraction leaves internal runtime space for weights, graphs, kernels, and transient allocations; it is not a video-workload reserve.", "acceptance_evidence": [ "startup health", "both-card memory use", "endpoint health after requests", "host-memory ownership for native KV offload", "clean mmap lifecycle" ] }, "control_observation": { "observed_status": "healthy at final campaign observation", "direct_endpoint": "http://127.0.0.1:39070/v1", "served_model": "deepseek-v4-flash-0731-r33-b12x-nospec-maxseq1-tp2-131k", "image": "voipmonitor/vllm@sha256:fdde59fed7f9fc12f9fd5ef1b3b3ea8d5097bf10ebad54b348497102c3a83f82", "runtime_version": "0.11.2.dev280+gilded.gnosis.v20.vllmfa13d33.b12x06db0f4.fi1ac6942.cu132.20260809.r33", "max_model_len": 131072, "max_num_seqs": 1, "max_num_batched_tokens": 8192, "speculative_decoding": "disabled", "weight_kernel_path": "B12X W4A8 with NVFP4 MoE weights and FP8 activations/dense path", "kv_cache": "FP8 DS-MLA", "router_observation": "direct-only qualification; no route or promotion changed during the campaign" }, "local_gates": [ { "id": "cache-integrity", "status": "passed", "result": "exact checkpoint snapshot complete with zero incomplete files or broken symlinks" }, { "id": "cold-start", "status": "passed", "result": "managed mode entry reached health in 323.5 seconds" }, { "id": "reasoning-prefix-fingerprint", "status": "passed", "result": { "low_prompt_tokens": 6, "high_prompt_tokens": 85, "max_prompt_tokens": 98 } }, { "id": "protocol-preflight", "status": "passed", "result": "6 of 6 checks passed: smoke, JSON, 20 of 20 tool calls, streaming tool call, tool-result continuation, and Responses API" }, { "id": "repeated-quality-high", "status": "passed-with-harness-caveat", "result": { "intelligence": "2 checks, 3 of 3 attempts each", "session_recall": "3 of 3 attempts", "tool_calling": "3 of 3 attempts", "context_32768_target": "passed with 23910 observed prompt tokens", "original_context_120000_target": "HTTP 400 Bad Request", "post_failure_short_coding_smoke": "passed" }, "decision": "Retain the HTTP 400 as a real failed attempt. Subsequent capacity requests and coding smokes passed, so it is not classified as a persistent endpoint or model failure." }, { "id": "near-limit-context-capacity", "status": "passed", "result": { "largest_observed_prompt_tokens": 119503, "requested_target_tokens": 119500, "ttft_ms": 17444.79, "end_to_end_ms": 17885.79, "effective_prefill_tokens_per_second": 7537.34, "decode_tokens_per_second": 73.86, "post_run_short_coding_smoke": "passed" }, "claim_limit": "This proves a single-request 119503-token capacity point on the 131072-token control, not 300K capacity or concurrency." }, { "id": "context-target-calibration", "status": "harness-caveat", "result": { "target_117500_actual_prompt_tokens": 87207, "target_118500_actual_prompt_tokens": 118830, "target_119500_actual_prompt_tokens": 119503, "target_120000_replay_clamped_to": 119808, "target_120000_replay_actual_prompt_tokens": 88939 }, "decision": "Use API-reported prompt tokens, never the requested target label, for context claims. Track deterministic calibration as a harness defect." }, { "id": "gpu-kv-capacity", "status": "measured", "result": { "available_gpu_kv_gib": 15.27, "gpu_kv_cache_tokens": 283917, "configured_context_tokens": 131072, "reported_concurrency_at_configured_context": 2.17 } }, { "id": "no-offload-over-300k", "status": "blocked-before-load", "result": "The measured 283917-token GPU KV pool is below the requested 300000-token floor and below 393216; removing a video reserve cannot change this result because that reserve gate is already disabled." }, { "id": "native-offload-absent-container-cleanup", "status": "passed", "result": "The managed absent-container path dry-ran and then reclaimed exactly one unmapped 4096-byte vLLM mmap fixture through the two-scan guard; the postcondition was clean and the control remained healthy." } ], "next_candidate": { "status": "translated-not-loaded", "reason_not_loaded": "The 131K r33 control was retained for continued operator testing.", "public_recipe": "configs/deepseek-v4-flash-0731-r33-b12x-nospec-maxseq1-393k-recipe.toml", "served_model": "deepseek-v4-flash-0731-r33-b12x-nospec-maxseq1-tp2-393k", "max_model_len": 393216, "max_num_seqs": 1, "max_num_batched_tokens": 4096, "speculative_decoding": "disabled", "kv_cache": "FP8 DS-MLA", "native_host_kv_offload_gib": 16, "native_filesystem_l2": "disabled", "quality_rationale": "Retains the released checkpoint and FP8 KV; adds host capacity instead of moving the cache to NVFP4 before a controlled quality comparison." }, "lifecycle_observations": { "rollback_attempt": "A pre-request mode-leave attempt failed during final router readmission after the transaction had run for about 543 seconds. Inspection showed the r33 control still healthy and the prior split owners absent.", "decision": "No route or promotion change. A 393K load requires a later explicitly reviewed lifecycle transaction." }, "promotion_state": "no-promotion", "raw_evidence": { "retention": "private operator evidence store", "public_record_kind": "sanitized decision summary with content hashes", "sha256": { "r33_131k_preflight_json": "F2E97D285EFB6BB2ED74620EA76006A02B798F0EDF5DF5F4A4831B5ED2F69511", "r33_131k_quality_high_json": "F83B95DFF27495E3E745C30E2F40724E6169E12B89385AC1EFBB1A385544E774", "r33_131k_reasoning_fingerprint_json": "88DEC7E0F0013984F5955264D3CCFF52300A425DD72C4F0F3179C3C02F756505", "r33_131k_context_boundary_high_json": "D5E8A20BD4683760C97BC3D401BA1FB1E5EDB36F225943E070D2C79BD0856514", "r33_131k_context_edge_high_json": "77CC5C7F325B8ABD8FC66C9F29AD04965CCE91FA84FE840F2DF7A11C98F8F58C", "r33_131k_context_120k_replay_high_json": "969712F2A6B7979EE6DB81C8F80724E67F11A1B042ACB4B7E32129D979F4307F", "native_offload_absent_acceptance_json": "66D1C9F7632444EFDDCF893017BB187554065492FDA5FF98E483D9C34F33DF01" }, "claim_limit": "This record does not claim that the 393216-token candidate has loaded or exercised a prompt above 300000 tokens." } }