{ "campaign_id": "2026-10-15-replacement-full", "configurations": [ { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "glm-r11", "model": { "repository": "brandonmusic/GLM-5.3-Flash-tr3-4bpw", "revision": "a5fee929cf4888b1824323e33e8a19b60129e025", "served_name": "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec" }, "qualification_role": "retained control configuration", "runtime": { "benchmark_native_generation_policy": { "request_temperature": 1.0, "request_top_p": 0.95, "thinking_path": "rendered Max" }, "context_tokens": 327680, "decode_context_parallel_size": 1, "engine": "vLLM EXL3", "engine_revision": "6dc2f516688fe6f84c6994dcd20fddf296853a6c", "execution_dtype": "bfloat16", "expert_parallel": true, "gpu_memory_utilization": 0.97, "image": "sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692", "kv_cache_dtype": "fp8_ds_mla", "max_num_batched_tokens": 2048, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 1, "configured_memory_swap_limit_mib": 53249, "configured_ram_mib": 53248, "host_reserve_mib": 12288, "readback_boundary": "configured contract only; the registry startup gate requires memory.max=55834574848 and memory.swap.max=1048576, but this supplement does not claim a separate retained live cgroup readback." }, "model_artifact_quantization": "uniform K4 EXL3/TR3 routed experts; native-precision non-routed tensors", "multimodal_limits": { "image": 8, "video": 0 }, "offload": "none recorded", "reasoning_and_tools": { "automatic_tool_choice": true, "reasoning_parser": "glm45", "tool_call_parser": "glm47" }, "relevant_serving_flags": [ "bfloat16 dtype", "safetensors load format", "b12x MoE backend", "FLASHINFER_MLA_SPARSE_SM120 attention backend", "chunked prefill enabled", "prefix caching enabled", "custom all-reduce disabled" ], "sampling": "generation configuration from the pinned checkpoint; no explicit sampling override recorded", "server_generation_defaults": "pinned checkpoint generation configuration; no explicit managed server sampling override recorded", "speculation": false, "tensor_parallel_size": 2 }, "source_binding": { "managed_registry_sha256": "3beba851889013e16084ccb5d6154d9dd77c4b742a25620a11bd68686f6afab6", "original_source_sha256": "f75795a7edf6efbd127851eade3d954ce04142102fad7b6698c5c9c1904045d9", "source_basename": "glm-agentic18-artifact.json" } }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "mimo-r4", "model": { "repository": "XiaomiMiMo/MiMo-V2.6-Flash-MOPD", "revision": "b12f5573350bb1704513fcf119fca48e498b0423", "served_name": "mimo26-mopd-fp8kv-omni-tp2-c4-327k-r4" }, "qualification_role": "unqualified r4 configuration; retained diagnostic evidence", "runtime": { "benchmark_native_generation_policy": { "request_temperature": 1.0, "request_top_p": 0.95, "thinking_path": "automatic" }, "context_tokens": 327680, "decode_context_parallel_size": "unknown", "engine": "vLLM", "engine_revision": "eb87980585c10f7e5ee8f498fd3830898101d208", "execution_dtype": "bfloat16", "expert_parallel": "unknown", "gpu_memory_utilization": 0.945, "image": "sha256:42ff601ea51f08c7598322ac22a363b843e1c8202c8358752bb50a01553b805b", "kv_cache_dtype": "fp8", "max_num_batched_tokens": 4096, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 53248, "configured_ram_mib": 53248, "evidence_boundary": "Configured managed recipe RAM/swap contract; this row does not claim a separate retained live cgroup measurement.", "host_reserve_mib": 12288 }, "model_artifact_quantization": "not recorded in retained managed registry or accepted native identity", "multimodal_limits": { "audio": 0, "image": 8, "video": 0 }, "offload": "no CPU KV offload recorded", "reasoning_and_tools": { "automatic_tool_choice": true, "reasoning_parser": "mimo", "tool_call_parser": "mimo" }, "relevant_serving_flags": [ "trust remote code", "MP distributed executor", "Marlin MoE backend", "full-decode-only CUDA graph mode", "custom all-reduce disabled" ], "server_generation_defaults": "generation configuration auto; no explicit managed server sampling override", "speculation": false, "tensor_parallel_size": 2 }, "source_binding": { "managed_registry_sha256": "b7b7a92d096af019fd61281c651564de7510ee81b5f60da44ac2e3018b699645", "original_source_sha256": "e79b69add3245c4a5b0b2e26e28ae23163e8563f1daf4fc80865327ed5743c7c", "recipe_snapshot_sha256": "615110a08629cbe919d6ad0b5a5a3e72fc5b79f4208361686c5a08264f147870", "source_basename": "mimo-agentic18-r1-native.json" } }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "mimo-r5-greedy", "model": { "repository": "XiaomiMiMo/MiMo-V2.6-Flash-MOPD", "revision": "b12f5573350bb1704513fcf119fca48e498b0423", "served_name": "mimo26-mopd-fp8kv-omni-tp2-c4-327k-r5-greedy" }, "qualification_role": "unqualified r5 configuration; sampling variant of r4", "runtime": { "benchmark_native_generation_policy": { "request_temperature": 0.0, "request_top_p": 0.95, "thinking_path": "automatic" }, "context_tokens": 327680, "decode_context_parallel_size": "unknown", "engine": "vLLM", "engine_revision": "eb87980585c10f7e5ee8f498fd3830898101d208", "execution_dtype": "bfloat16", "expert_parallel": "unknown", "gpu_memory_utilization": 0.945, "image": "sha256:42ff601ea51f08c7598322ac22a363b843e1c8202c8358752bb50a01553b805b", "kv_cache_dtype": "fp8", "max_num_batched_tokens": 4096, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 53248, "configured_ram_mib": 53248, "evidence_boundary": "Configured managed recipe RAM/swap contract; this row does not claim a separate retained live cgroup measurement.", "host_reserve_mib": 12288 }, "model_artifact_quantization": "not recorded in retained managed registry or accepted native identity", "multimodal_limits": { "audio": 0, "image": 8, "video": 0 }, "offload": "no CPU KV offload recorded", "reasoning_and_tools": { "automatic_tool_choice": true, "reasoning_parser": "mimo", "tool_call_parser": "mimo" }, "relevant_serving_flags": [ "trust remote code", "MP distributed executor", "Marlin MoE backend", "full-decode-only CUDA graph mode", "custom all-reduce disabled" ], "server_generation_defaults": { "effective_greedy_normalization": "source-inferred only; not a response echo", "override_generation_config": { "temperature": 0.0, "top_p": 0.95 } }, "speculation": false, "tensor_parallel_size": 2 }, "source_binding": { "managed_registry_sha256": "6fadf1b1722cbe1def1142bafdf0bc480e46cd82cdf824b3e0345c6267daf079", "original_source_sha256": "e6fda11ac6ee2d057d436251417edcd00b8835178af9056c7a36b178eea0ae0b", "source_basename": "mimo-r5-greedy-agentic18-r1-native.json" } }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "mistral-r2", "model": { "repository": "mistralai/Mistral-Medium-3.5-128B", "revision": "22b2b868a15677cfa6061277ed2f653d1349a9ab", "served_name": "mistral35-fp8kv-tp2-c4-262k-r2" }, "qualification_role": "rejected resource-policy configuration", "runtime": { "benchmark_native_generation_policy": { "evidence_status": "Declared source-supported comparison policy; no r2 agentic benchmark executed. Bound r2 protocol preflight uses its separate fixed protocol sampling path.", "reasoning_effort": "high", "request_temperature": 0.7, "request_top_p": 0.95, "thinking_path": "default" }, "context_tokens": 262144, "decode_context_parallel_size": "unknown", "engine": "vLLM", "engine_revision": "3f1bcfd82bb2e93fcd364fc7a5491718453a77e2", "execution_dtype": "bfloat16", "expert_parallel": "not applicable", "gpu_memory_utilization": 0.95, "image": "vllm/vllm-openai@sha256:cb9bc64f0f038c2c2ebf13b31248fb87aca40c356c3d20f363be087cfcff1fcd", "kv_cache_dtype": "fp8", "max_num_batched_tokens": 4096, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 53248, "configured_ram_mib": 53248, "evidence_boundary": "Configured managed recipe RAM/swap contract; this row does not claim a separate retained live cgroup measurement.", "host_reserve_mib": 12288 }, "model_artifact_quantization": "not recorded in retained managed registry or accepted native identity", "multimodal_limits": { "image": 8, "video": 0 }, "offload": "no host KV offload recorded", "reasoning_and_tools": { "automatic_tool_choice": true, "native_reasoning_controls": [ "none", "high" ], "reasoning_parser": "mistral", "tool_call_parser": "mistral" }, "relevant_serving_flags": [ "MP distributed executor" ], "server_generation_defaults": "generation configuration auto; no explicit managed server sampling override", "speculation": false, "tensor_parallel_size": 2 }, "source_binding": { "managed_registry_sha256": "b7b7a92d096af019fd61281c651564de7510ee81b5f60da44ac2e3018b699645", "original_source_sha256": "fac143fa58a36fe706c71a6ac092d759abddc6dc3316f00aec558c13a47b286b", "recipe_snapshot_sha256": "21937aaec6d37686c18ecbc96638ec8fe00eece552227d618947d8507fb33cd3", "source_basename": "mistral-r2-model-aware-preflight.json" } }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "mistral-r3", "model": { "repository": "mistralai/Mistral-Medium-3.5-128B", "revision": "22b2b868a15677cfa6061277ed2f653d1349a9ab", "served_name": "mistral35-fp8kv-tp2-c4-262k-r3" }, "qualification_role": "rejected reliability configuration", "runtime": { "benchmark_native_generation_policy": { "evidence_status": "Executed by the retained r3 agentic round; separate from deterministic protocol preflight.", "reasoning_effort": "high", "request_temperature": 0.7, "request_top_p": 0.95, "thinking_path": "default" }, "context_tokens": 262144, "decode_context_parallel_size": "unknown", "engine": "vLLM", "engine_revision": "3f1bcfd82bb2e93fcd364fc7a5491718453a77e2", "execution_dtype": "bfloat16", "expert_parallel": "not applicable", "gpu_memory_utilization": 0.94, "image": "vllm/vllm-openai@sha256:cb9bc64f0f038c2c2ebf13b31248fb87aca40c356c3d20f363be087cfcff1fcd", "kv_cache_dtype": "fp8", "max_num_batched_tokens": 4096, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 53248, "configured_ram_mib": 53248, "evidence_boundary": "Configured managed recipe RAM/swap contract; this row does not claim a separate retained live cgroup measurement.", "host_reserve_mib": 12288 }, "model_artifact_quantization": "not recorded in retained managed registry or accepted native identity", "multimodal_limits": { "image": 8, "video": 0 }, "offload": "no host KV offload recorded", "reasoning_and_tools": { "automatic_tool_choice": true, "native_reasoning_controls": [ "none", "high" ], "reasoning_parser": "mistral", "tool_call_parser": "mistral" }, "relevant_serving_flags": [ "MP distributed executor" ], "server_generation_defaults": "generation configuration auto; no explicit managed server sampling override", "speculation": false, "tensor_parallel_size": 2 }, "source_binding": { "managed_registry_sha256": "a8495d2db46959352b91ec509b64fb19e470aded43bd6be78ae35413c0843879", "original_source_sha256": "5f23ba96489b23e819c5a9d3b415d9ed154bf3a7d66c2e575e6a5fb495888230", "source_basename": "mistral-r3-agentic18-r1-evidence.json" } }, { "hardware": { "configuration": "one RTX PRO 6000 Blackwell Max-Q compute GPU", "vram_mib_per_gpu": 97887 }, "id": "deepseek-runtime-r4", "model": { "quantization_candidate": { "evidence_status": "external-prior metadata only; no model load, serving configuration, or numerical qualification claim", "format": "EXL3 2.7bpw", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "repository": "deepseek-ai/DeepSeek-V4.1-Flash", "revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "served_name": "not applicable: Stage 2 numerical-test contract" }, "qualification_role": "unqualified runtime build and independently reviewed Stage 2 r6 one-GPU numerical-test contract; not a serving configuration", "runtime": { "base_image": "vllm/vllm-openai@sha256:c1c9f6fd5c109ba7f0546a59f5b2f15fb87f64c77782e90a27b648b42a8e67c3", "build_precision_and_kv": "not applicable: r4 is a build and Stage 2 numerical-test contract, not a serving recipe", "build_resource_contract": { "cpus": 2, "effective_swap_mib": 0, "max_jobs": 1, "nvcc_threads": 1, "ram_mib": 16384, "timeout_seconds": 1800 }, "context_tokens": "not applicable", "decode_context_parallel_size": "not applicable", "engine": "vLLM 0.31.0+dsv41.sm120 with EXL3 plugin", "gpu_memory_utilization": "not applicable", "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "max_num_seqs": "not applicable", "multimodal_limits": "not applicable", "offload": "none recorded", "relevant_runtime_controls": { "cuda": "13.0.2", "exl3_int8_gemv": false, "rdma_all_reduce": false, "target_architecture": "SM120", "torch": "2.13.0+cu130", "vllm_plugins": "exl3_moe" }, "source_pins": { "exllamav3": "151539c77abc7ab7425d30da7a4e8e3c5c154e7b", "recipe_plugin": "cc59c3609e8e671418f39eb92aa16b0e4f770b77", "vllm": "6e517b15c1833cf72a7f557ee32524d98682e617" }, "speculation": false, "stage2_resource_contract": { "compute_gpus": 1, "cpus": 8, "effective_swap_mib": 0, "host_mounts": "none", "network": "none", "ram_mib": 24576, "required_compute_capability": [ 12, 0 ], "root_filesystem": "read-only", "timeout_seconds": 2700, "tmpfs": "bounded 4 GiB /tmp for JIT" }, "tensor_parallel_size": "not applicable: Stage 2 uses exactly one compute GPU" }, "source_binding": { "additional_stage2_original_source_sha256": "1252b3adc0347cbc7d85b9ae745e20e76ee75bf06a566dd294c00531ee6ff2f8", "additional_stage2_source_basename": "deepseek-stage2-r4-gpu-test-r2.json", "build_config_sha256": "22a5e8ed63daa8785aead38e67cccb8fc92c0ca10373f2eaf967b731ffbed75a", "original_source_sha256": "67a9203250e03a5be7a79c4bef33ee4a98596dc5fe2dda1cd5df378fdc718f86", "revision_source_basename": "source-registry.json", "revision_source_sha256": "3e27c3b852dfc1656a84ef5f98b96b7338e178f41c3af8b5be26ff58107353b3", "source_basename": "deepseek-build-r4.json", "source_lock_sha256": "22b2b086d0d5bdb144c77133b2e1b4fcb7dc22412dd57158794953cf394f8ecf", "stage2_r6_native_source_basename": "deepseek-stage2-r6.native.json", "stage2_r6_native_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "stage2_r6_review_source_basename": "independent-stage2-r6-review.json", "stage2_r6_review_source_sha256": "95a190fd0dc0c3ff58408ca39fc0e652e878e949105161f65acd522e097ab555" }, "status": "Verified Stage 2 r6 numerical runtime compatibility only; no model weights were loaded. This row does not describe a Stage 3 serving plan, capacity, quality, or promotion result." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage3-r8", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage3-dummy-r8", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "unqualified Stage 3 dummy hardware-proxy configuration", "runtime": { "context_tokens": 65536, "cuda_graph_mode": "FULL_DECODE_ONLY", "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "dummy", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "EXL3 2.7bpw candidate artifact with dummy parameter allocation and verified official Engram disk tables; no main weight files" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "The helper-produced dummy EXL3 object omitted the original candidate dense, head, cooperative, and fused controls; r8 is not a faithful real-quant residency proxy." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "hardware_source_basename": "deepseek-stage2-r6.native.json", "hardware_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "original_source_sha256": "9961d1d5149bd19ebb0ea0a947d72b949cdb2254a180d920800e2be215eeb70c", "source_basename": "serve-recipes.stage3-dummy-r8-engram.toml" }, "status": "Stage 3 dummy hardware-proxy configuration only; no quality, throughput, real main-weight fit, or promotion conclusion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage3-r9", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage3-dummy-r9", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "unqualified Stage 3 dummy hardware-proxy configuration", "runtime": { "context_tokens": 65536, "cuda_graph_mode": "FULL_DECODE_ONLY", "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "dummy", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "EXL3 2.7bpw candidate artifact with dummy parameter allocation and verified official Engram disk tables; no main weight files" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "After helper generation, r9 restored and asserted exact equality of the verified candidate exl3_moe object, including head K4 and dense map." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "delta_source_basename": "delta-r9.md", "delta_source_sha256": "6c95c4491cdc77fbad4a122057462e2dabe4a26024024f45e4cd706eb1587056", "hardware_source_basename": "deepseek-stage2-r6.native.json", "hardware_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "original_source_sha256": "2d5cf0c1bc3c2bff0ea26127f5862ace04a384e64e8bb6efefa9e88b73d74ea9", "source_basename": "serve-recipes.stage3-dummy-r9-engram.toml" }, "status": "Stage 3 dummy hardware-proxy configuration only; no quality, throughput, real main-weight fit, or promotion conclusion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage3-r10", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage3-dummy-r10", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "unqualified Stage 3 dummy hardware-proxy configuration", "runtime": { "context_tokens": 65536, "cuda_graph_mode": "FULL_DECODE_ONLY", "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "dummy", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "EXL3 2.7bpw candidate artifact with dummy parameter allocation and verified official Engram disk tables; no main weight files" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "r10 preserves r9 restored candidate controls and adds only VLLM_EXL3_DENSE_MROW=0 as a supported diagnostic dense-path selection; no runtime outcome is asserted." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "delta_source_basename": "delta-r10.md", "delta_source_sha256": "48322b9ec4c4c76725f10ce26d62797e61c23667401c9dfd2d5b163713f612c0", "hardware_source_basename": "deepseek-stage2-r6.native.json", "hardware_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "original_source_sha256": "8b44e91096ec877711316e39861e4c233ae311ff7ed4240881ab49700fcccbe6", "source_basename": "serve-recipes.stage3-dummy-r10-engram.toml" }, "status": "Stage 3 dummy hardware-proxy configuration only; no quality, throughput, real main-weight fit, or promotion conclusion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage3-r11", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage3-dummy-r11", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "unqualified Stage 3 dummy hardware-proxy configuration", "runtime": { "context_tokens": 65536, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "dummy", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "EXL3 2.7bpw candidate artifact with dummy parameter allocation and verified official Engram disk tables; no main weight files" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "r11 preserves r10 MROW=0 and adds only --disable-custom-all-reduce, the pinned NCCL-fallback diagnostic transport selection; no runtime outcome is asserted by this configuration row." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "delta_source_basename": "delta-r11.md", "delta_source_sha256": "f5ea8256a0286f019bbc591be3d8f0596729c807665db5a24bb7f960d6578f9d", "hardware_source_basename": "deepseek-stage2-r6.native.json", "hardware_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "original_source_sha256": "9adc74db303105c48ecc1e76efa5355ab3db0cd138bf3107d6a6ac55d39cd676", "source_basename": "serve-recipes.stage3-dummy-r11-engram.toml", "source_manifest_basename": "manifest.json", "source_manifest_sha256": "1d5113ca2f315ca22049b7176fb888e76b524f8e336280335e0bf6fda9f654c1" }, "status": "Stage 3 dummy hardware-proxy configuration only; no quality, throughput, real main-weight fit, or promotion conclusion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage3-r12", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage3-dummy-r12", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "unqualified Stage 3 dummy hardware-proxy configuration", "runtime": { "context_tokens": 65536, "cpu_offload_gb_per_tp_worker": 4, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "dummy", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "EXL3 2.7bpw candidate artifact with dummy parameter allocation and verified official Engram disk tables; no main weight files" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "r12 preserves r11 MROW=0 and disabled custom all-reduce, and adds only --cpu-offload-gb 4 per TP worker as a bounded generic-UVA diagnostic; the configuration row does not establish EXL3 offload support, real-weight fit, quality, throughput, or promotion." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "delta_source_basename": "delta-r12.md", "delta_source_sha256": "bc61d2070282b8589d63a3250241c481309fd28b5d7f1b42e4b6bb8be0e082e3", "hardware_source_basename": "deepseek-stage2-r6.native.json", "hardware_source_sha256": "0b9d0e80dbcaa300835c900dd401b44d90915669c087c209feaf812dbce1a782", "offline_proof_source_basename": "offline-proof.json", "offline_proof_source_sha256": "9679c2c2332e8f3432bb44689911c1100ee52f1958cd7a96c4cb087e3b9380bd", "original_source_sha256": "46973d9517786f52307e7ac6f80160d0ad9d8fff563b42c57e2574ce483d4aff", "preexecution_review_source_basename": "independent-review-r2.json", "preexecution_review_source_sha256": "c39820a5d40812f34938d7cc0b08ac2068ce906877158386e9e4f6a75275d135", "root_terminal_record_source_basename": "root-terminal-record.json", "root_terminal_record_source_sha256": "2555576b40005be026b18335735286e811770235b87b8092ca42bfb30b25120f", "source_basename": "serve-recipes.stage3-dummy-r12-engram.toml", "source_manifest_basename": "manifest.json", "source_manifest_sha256": "1d5113ca2f315ca22049b7176fb888e76b524f8e336280335e0bf6fda9f654c1", "terminal_review_source_basename": "independent-terminal-review.json", "terminal_review_source_sha256": "0ba617460d27a4e68ff38d6ca86890c93b706428e53e3d6aef673d082b6bc39a" }, "status": "Stage 3 dummy hardware-proxy configuration with accepted bounded dummy-path evidence only; no quality, throughput, real main-weight fit, or promotion conclusion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage4-r7", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage4-real-text-r7", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "attempted bounded real-weight 64K/C1/image0 feasibility; pre-engine integrity rejection, no correctness request", "runtime": { "context_tokens": 65536, "cpu_offload_gb_per_tp_worker": 4, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "dense_mrow": 0, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "engram_config": { "cpu_offload": true, "disk_offload": true, "disk_threads": 256, "dp_shared_memory": false }, "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "safetensors", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "Original EXL3 2.7bpw candidate main weights and verified official Engram disk tables" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "Pinned real main pack; exact original candidate plan and index; no dummy weights. r8 corrects the r7 index checksum transposition only, apart from attempt identity." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "original_source_sha256": "a8d0707d4568bbcdf69fdf775b2581c55de88d19a3c62d315765526bc887b037", "public_native_derivative": "native/deepseek-stage4.native.json", "public_native_derivative_sha256": "519aaa9478f26a875eeea7fec1ee1ba41dc182b298e73e50808cec3a4d7ed438", "source_basename": "serve-recipes.stage4-real-text-r7-c1-64k.toml", "terminal_record_source_basename": "deepseek-stage4-r7-terminal.json", "terminal_record_source_sha256": "0ec2f78cc2cf295c9814fda5d04b1b5c4499f177ad34caf1021fb48c4bf5b324" }, "status": "Pre-engine checksum failure; no GPU weight load or requests." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-stage4-r8", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-stage4-real-text-r8", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "bounded real-weight 64K/C1/image0 feasibility and basic correctness", "runtime": { "context_tokens": 65536, "cpu_offload_gb_per_tp_worker": 4, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "dense_mrow": 0, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "engram_config": { "cpu_offload": true, "disk_offload": true, "disk_threads": 256, "dp_shared_memory": false }, "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 536870912, "load_format": "safetensors", "max_num_batched_tokens": 2048, "max_num_seqs": 1, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "Original EXL3 2.7bpw candidate main weights and verified official Engram disk tables" }, "multimodal_limits": { "image": 0 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "Pinned real main pack; exact original candidate plan and index; no dummy weights. r8 corrects the r7 index checksum transposition only, apart from attempt identity." }, "source_binding": { "candidate_config_source_basename": "artifact-config.json", "candidate_config_source_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "original_source_sha256": "0c711e9a858a522c68e4dad0768e85e11a8f5779c1424aaadbd78a36d888d08f", "public_native_derivative": "native/deepseek-stage4.native.json", "public_native_derivative_sha256": "519aaa9478f26a875eeea7fec1ee1ba41dc182b298e73e50808cec3a4d7ed438", "source_basename": "serve-recipes.stage4-real-text-r8-c1-64k.toml", "terminal_record_source_basename": "deepseek-stage4-r8-terminal.json", "terminal_record_source_sha256": "42da253b1f8238b3bfdcd46cb0ab84b79bc6412a1cb1be2d55345f3dc254664c" }, "status": "Bounded real-weight basic 2/2 and resource gate accepted independently; no native-quality, full-role or promotion result." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-full-role-r2", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-full-role-r2", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "full-role admission and initial nonthinking correctness/resource gate", "runtime": { "context_tokens": 327680, "cpu_offload_gb_per_tp_worker": 4, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "dense_mrow": 0, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "engram_config": { "cpu_offload": true, "disk_offload": true, "disk_threads": 256, "dp_shared_memory": false }, "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 2147483648, "load_format": "safetensors", "max_num_batched_tokens": 2048, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "Original EXL3 2.7bpw candidate main weights and verified official Engram disk tables" }, "multimodal_limits": { "image": 8 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "Full-role327680/C4/image8 with2GiB fixedKV; original weights and allother r8controls retained." }, "source_binding": { "original_source_sha256": "3a9bd964cec79f84f96433598aae4e96841a256db5d42c2e53bf6db9aa159dac", "public_native_derivative": "native/deepseek-full-role-r2.native.json", "public_native_derivative_sha256": "b2c0faa448a8325168754480fa9766d9197179e03e1327968129ac0bb4827a4f", "source_basename": "serve-recipes.deepseek-full-role-r2.toml", "terminal_record_source_basename": "deepseek-full-role-r2-terminal.json", "terminal_record_source_sha256": "7f91a2d2e394c0ca02d47501648572ffde1b7a37b80286b826df5ea47fc8b5d5" }, "status": "Startup and basic2/2 correctness passed; activeGPUreserve4087MiB/card failed4096policy. Unloaded, nofollowon qualityrequests orpromotion." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-full-role-r3", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-full-role-r3", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "full-role resource feasibility and bounded protocol controls; resource policy rejection", "runtime": { "context_tokens": 327680, "cpu_offload_gb_per_tp_worker": 4, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "dense_mrow": 0, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "engram_config": { "cpu_offload": true, "disk_offload": true, "disk_threads": 256, "dp_shared_memory": false }, "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 1610612736, "load_format": "safetensors", "max_num_batched_tokens": 2048, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "Original EXL3 2.7bpw candidate main weights and verified official Engram disk tables" }, "multimodal_limits": { "image": 8 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "Exact real main pack and official Engram; full role with 1.5 GiB KV per rank, CPU offload 4 GiB per worker; all runtime pins retained." }, "source_binding": { "original_source_sha256": "ff7897d7926d0aa340e86dc40c1233c820eb4f9f003659b7af9186a3657816f0", "public_native_derivative": "native/deepseek-full-role-r3.native.json", "public_native_derivative_sha256": "e71b41d4269a809dfef6b82e1e20d2070fa573b0d12cb4d0a53633058853599a", "source_basename": "serve-recipes.deepseek-full-role-r3.toml", "terminal_record_source_basename": "deepseek-full-role-r3-terminal.json", "terminal_record_source_sha256": "64742e233fd235005268876e48fbfce5154ab87638d0bf4f7abfee9e292465b3" }, "status": "Startup and basic2/2/model-aware5/5 passed, but observed active reserve3961MiB/card failed4096floor. Unloaded; no native-quality qualification." }, { "hardware": { "configuration": "two RTX PRO 6000 Blackwell Max-Q GPUs", "vram_mib_per_gpu": 97887 }, "id": "deepseek-full-role-r4", "model": { "official_engram_repository": "deepseek-ai/DeepSeek-V4.1-Flash", "official_engram_revision": "2cba9e42aa026125f3ed06c6d98c1db82f7ca027", "recipe_identity": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw-full-role-r4", "repository": "sudoingx/DeepSeek-V4.1-Flash-EXL3-2.7bpw", "revision": "f7451e9617f675087881b5016fb210a526c7f99c" }, "qualification_role": "full-role candidate: protocol and resource gates passed; native agentic reliability gate failed", "runtime": { "context_tokens": 327680, "cpu_offload_gb_per_tp_worker": 6, "cuda_graph_mode": "FULL_DECODE_ONLY", "custom_all_reduce_disabled": true, "dense_mrow": 0, "engine": "vLLM EXL3 DeepSeek V4.1 SM120 control runtime", "engram_config": { "cpu_offload": true, "disk_offload": true, "disk_threads": 256, "dp_shared_memory": false }, "flashinfer_autotune": false, "gpu_memory_utilization": 0.95, "image": "sha256:b6db406afb24809ed7b27c7312556487cb7099baf73fd7423183e9d6f7061e4b", "kv_cache_memory_bytes": 1610612736, "load_format": "safetensors", "max_num_batched_tokens": 2048, "max_num_seqs": 4, "memory_contract": { "configured_effective_swap_mib": 0, "configured_memory_swap_limit_mib": 65536, "configured_ram_mib": 65536, "evidence_boundary": "Configured managed contract only; values do not establish capacity, quality, or promotion.", "host_reserve_mib": 12288 }, "model_artifact_quantization": { "candidate_configuration_sha256": "d1758fb479f7b360fd6004a2a5d1dc2fdffe70b11c2a10a6e33740a80c9ec253", "candidate_exl3_moe": { "coop_max_tokens": 32, "decode_kernel": "coop", "dense_map": { "drafter": { "attn.wo_b": 5, "attn.wq_a": 5, "attn.wq_b": 5, "ffn.shared_experts.w1": 5, "ffn.shared_experts.w2": 5 }, "k": { "attn.wo_a": 5, "attn.wo_b": 4, "attn.wq_a": 4, "attn.wq_b": 4, "ffn.shared_experts.w1": 4, "ffn.shared_experts.w2": 4 }, "layers": [ 0, 39 ] }, "exllamav3_version": "1.6.0", "fused_rows": 256, "head": "exl3", "head_k": 4, "plan_file": "exl3_moe_plan.json", "tp_layout": "channels" }, "format": "Original EXL3 2.7bpw candidate main weights and verified official Engram disk tables" }, "multimodal_limits": { "image": 8 }, "rdma_all_reduce": false, "speculation": false, "tensor_parallel_size": 2, "variant_detail": "Exact real main pack and official Engram; configured1.5GiB KV per rank and6GiB CPU offload per worker; full327680/C4/image8 role. CPU parameter accounting may overshoot a parameter boundary; not resident RAM." }, "source_binding": { "source_basename": "serve-recipes.deepseek-full-role-r4.toml", "original_source_sha256": "c567ca5443a3ee8d24415fc67531ccb0d3f9290f37b57e047c357cf13664c32a", "terminal_record_source_basename": "deepseek-full-role-r4-oracle8-terminal-gate.json", "terminal_record_source_sha256": "dff2a904acfa3b930b744434ea378aff67d9330e884aea68d0bee2c8945d5c45", "independent_terminal_review_sha256": "736b94db0a1376b617e881fa4c4e465e7c1fbeef58bd5dda780ac738c47e8b28", "public_native_derivative": "native/deepseek-full-role-r4.native.json", "public_native_derivative_sha256": "8db47000bfa1896eace7202b2e6d3e7a94031bb83dd84057d06887fd7aabef7e" }, "status": "Basic2/2 and model-aware5/5 passed; nativeOracle8 first round17/18 failed exact tool protocol. Resource minimum6539MiB/card passed. Unloaded; no follow-on rounds or finalist gates." } ], "observed_date": "2026-10-15", "redaction_policy": { "omitted": [ "private paths", "ports and endpoints", "physical GPU UUIDs and bus identities", "container names and IDs", "named volumes and mount paths" ], "unknown_value": "unknown means the retained evidence did not record the field; it is not an inferred default." }, "schema": "anvil-serving.replacement-publication-configuration-supplement/v1", "scope": "Public-safe reconstruction details for retained configurations. This supplement contains no endpoint, host, port, physical-device, container, volume, or private-path identity. Configured values are not measured capacity, quality, or promotion claims." }