{ "schemaVersion": "1.0.0", "datasetVersion": "2026.08.18.3", "generatedAt": "2026-08-19T03:21:00Z", "license": "CC-BY-4.0", "models": [ { "id": "qwen3.8-27b", "name": "Qwen3.8-27B", "publisher": "Qwen", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0", "parameters": 27781427952, "architecture": "Qwen3_5ForConditionalGeneration", "layers": 64, "fullAttentionLayers": 16, "kvHeads": 4, "headDimension": 256, "nativeContext": 262144, "kvBytesPerTokenFp16": 65536, "license": "Apache-2.0", "evidence": "reported", "sourceIds": [ "qwen38-hf-api", "qwen38-config" ] }, { "id": "qwen3.6-27b", "name": "Qwen3.6-27B", "publisher": "Qwen", "repository": "Qwen/Qwen3.6-27B", "revision": "unlocked-latest", "parameters": 27000000000, "architecture": "Qwen3_5ForConditionalGeneration", "layers": 64, "fullAttentionLayers": 16, "kvHeads": 4, "headDimension": 256, "nativeContext": 262144, "kvBytesPerTokenFp16": 65536, "license": "Apache-2.0", "evidence": "reported", "sourceIds": [ "qwen36-model-card" ] }, { "id": "qwen3-8b", "name": "Qwen3-8B", "publisher": "Qwen", "repository": "Qwen/Qwen3-8B", "revision": "unlocked-latest", "parameters": 8200000000, "architecture": "Qwen3ForCausalLM", "layers": 36, "fullAttentionLayers": 36, "kvHeads": 8, "headDimension": 128, "nativeContext": 32768, "kvBytesPerTokenFp16": 73728, "license": "Apache-2.0", "evidence": "reported", "sourceIds": [ "qwen3-8b-model-card" ] } ], "hardware": [ { "id": "rtx-4090-24gb", "name": "NVIDIA GeForce RTX 4090", "vendor": "NVIDIA", "memoryGb": 24, "memoryType": "GDDR6X", "topology": "discrete", "evidence": "reported", "sourceIds": [ "nvidia-rtx4090" ] }, { "id": "rtx-5090-32gb", "name": "NVIDIA GeForce RTX 5090", "vendor": "NVIDIA", "memoryGb": 32, "memoryType": "GDDR7", "topology": "discrete", "evidence": "reported", "sourceIds": [ "nvidia-rtx5090" ] }, { "id": "mac-studio-m4-max-64gb", "name": "Mac Studio M4 Max · 64GB", "vendor": "Apple", "memoryGb": 64, "memoryType": "Unified memory", "topology": "unified", "evidence": "reported", "sourceIds": [ "apple-mac-studio" ] }, { "id": "mac-studio-m3-ultra-512gb", "name": "Mac Studio M3 Ultra · 512GB", "vendor": "Apple", "memoryGb": 512, "memoryType": "Unified memory", "topology": "unified", "evidence": "reported", "sourceIds": [ "apple-mac-studio" ] } ], "quantizations": [ { "id": "q4-k-m", "name": "Q4_K_M", "estimatedBitsPerWeight": 4.8, "qualityLabel": "Highest fit efficiency" }, { "id": "q5-k-m", "name": "Q5_K_M", "estimatedBitsPerWeight": 5.7, "qualityLabel": "Balanced" }, { "id": "q6-k", "name": "Q6_K", "estimatedBitsPerWeight": 6.6, "qualityLabel": "Higher fidelity" }, { "id": "q8-0", "name": "Q8_0", "estimatedBitsPerWeight": 8.5, "qualityLabel": "Near-source precision" }, { "id": "bf16", "name": "BF16", "estimatedBitsPerWeight": 16, "qualityLabel": "Source precision" } ], "runs": [ { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-8192-1787105510", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "8192", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 8192, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "fixed" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 17418944512, "peakMemoryMiB": 16612, "loadSeconds": 5.673754267394543, "promptTokensPerSecond": 2036.8872931125297, "generationTokensPerSecond": 47.72116865203187, "timeToFirstTokenSeconds": 0.2587182894349098, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 2021.5420575507756, "median": 2036.0281544518232, "max": 2053.0916673349907 }, "generationTokensPerSecond": { "min": 47.69089689678714, "median": 47.72821225901011, "max": 47.74439680029835 }, "timeToFirstTokenSeconds": { "min": 0.25660601630806923, "median": 0.2586691938340664, "max": 0.26087965816259384 } } }, "evaluation": null, "timestamp": "2026-08-19T02:11:33.743106Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-fixed-8192.json", "serverLog": "logs/server-fixed-8192.log", "memoryLog": "logs/memory-fixed-8192.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/fixed-8192.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:aa3fb4e5afec81b14d282228a02aab9a09d826a0668045809bb8857680fcc053" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-32768-1787105527", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "32768", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 32768, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "fixed" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 19054723072, "peakMemoryMiB": 18172, "loadSeconds": 5.616523429751396, "promptTokensPerSecond": 2003.1333730964436, "generationTokensPerSecond": 47.73488996752966, "timeToFirstTokenSeconds": 0.262914693603913, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 1976.1244032065704, "median": 2006.7728583970902, "max": 2026.5028576856705 }, "generationTokensPerSecond": { "min": 47.7147633986437, "median": 47.7438583353258, "max": 47.746048168619495 }, "timeToFirstTokenSeconds": { "min": 0.259922631084919, "median": 0.2623755671083927, "max": 0.2664458826184273 } } }, "evaluation": null, "timestamp": "2026-08-19T02:11:50.679270Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-fixed-32768.json", "serverLog": "logs/server-fixed-32768.log", "memoryLog": "logs/memory-fixed-32768.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/fixed-32768.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:3296ccc37e85fb4c06de7c21aed0ea64d9a7921f1a4d03ec952f290477fb13fd" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-65536-1787105544", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "65536", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 65536, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "fixed" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 21235761152, "peakMemoryMiB": 20252, "loadSeconds": 5.633943635970354, "promptTokensPerSecond": 2008.9435371619509, "generationTokensPerSecond": 47.7455286578726, "timeToFirstTokenSeconds": 0.26218712081511814, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 1992.4582342617202, "median": 2015.9861400952868, "max": 2018.3862371288456 }, "generationTokensPerSecond": { "min": 47.735711047847715, "median": 47.748920385630804, "max": 47.75195454013927 }, "timeToFirstTokenSeconds": { "min": 0.2608868181705475, "median": 0.2613201439380646, "max": 0.2643544003367424 } } }, "evaluation": null, "timestamp": "2026-08-19T02:12:07.504115Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-fixed-65536.json", "serverLog": "logs/server-fixed-65536.log", "memoryLog": "logs/memory-fixed-65536.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/fixed-65536.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:04ce8988835134494cbcdb562ec8b33b4b4ff5571599473b46993bf4fd70eb8e" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-98304-1787105569", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "98304", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 98304, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 23416799232, "peakMemoryMiB": 22332, "loadSeconds": 6.174851767718792, "promptTokensPerSecond": 1943.471959622185, "generationTokensPerSecond": 47.590878601181004, "timeToFirstTokenSeconds": 0.27111514036854106, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 1883.5232184702995, "median": 1968.8672860395004, "max": 1978.0253743567555 }, "generationTokensPerSecond": { "min": 47.46128821894976, "median": 47.58570015988046, "max": 47.72564742471279 }, "timeToFirstTokenSeconds": { "min": 0.26637742295861244, "median": 0.2673618234694004, "max": 0.2796061746776104 } } }, "evaluation": null, "timestamp": "2026-08-19T02:12:31.911321Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-98304.json", "serverLog": "logs/server-boundary-98304.log", "memoryLog": "logs/memory-boundary-98304.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-98304.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:77dffb7c3fe17acbc3f76dd181aa48449f12b9e41885a2877dc56412093e1b42" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-114688-1787105586", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "114688", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 114688, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 24507318272, "peakMemoryMiB": 23372, "loadSeconds": 5.627897329628468, "promptTokensPerSecond": 1922.4098020737977, "generationTokensPerSecond": 47.651894902489154, "timeToFirstTokenSeconds": 0.2758141942322254, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 1746.0398656367759, "median": 1969.200473838864, "max": 2051.9890667457535 }, "generationTokensPerSecond": { "min": 47.59943315446691, "median": 47.63489019407655, "max": 47.72136135892401 }, "timeToFirstTokenSeconds": { "min": 0.25662970542907715, "median": 0.2671837843954563, "max": 0.3036290928721428 } } }, "evaluation": null, "timestamp": "2026-08-19T02:12:49.417423Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-114688.json", "serverLog": "logs/server-boundary-114688.log", "memoryLog": "logs/memory-boundary-114688.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-114688.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:10551f250f4a8dbbfeacf58e3eb9ca2610baa423e82669746c1c1317f6c9876e" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-122880-1787105603", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "122880", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 122880, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 25052577792, "peakMemoryMiB": 23892, "loadSeconds": 6.117317266762257, "promptTokensPerSecond": 2016.6466630803825, "generationTokensPerSecond": 47.69292339535946, "timeToFirstTokenSeconds": 0.2612360914548238, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 2000.1171943668573, "median": 2011.5981204130062, "max": 2038.2246744612835 }, "generationTokensPerSecond": { "min": 47.67968860282744, "median": 47.69161326082205, "max": 47.707468322428866 }, "timeToFirstTokenSeconds": { "min": 0.2585005536675453, "median": 0.2620650380849838, "max": 0.2631426826119423 } } }, "evaluation": null, "timestamp": "2026-08-19T02:13:06.290447Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-122880.json", "serverLog": "logs/server-boundary-122880.log", "memoryLog": "logs/memory-boundary-122880.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-122880.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:ef99b15f0269df376739bad22610372dba070dcaedca023b3e651ed17e2ada83" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-124928-1787105628", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "124928", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 124928, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 1, "measuredRuns": 3, "peakMemoryBytes": 25188892672, "peakMemoryMiB": 24022, "loadSeconds": 6.187585934996605, "promptTokensPerSecond": 1975.6633710088265, "generationTokensPerSecond": 47.71340208975224, "timeToFirstTokenSeconds": 0.266362043718497, "outcome": "success", "actualPromptTokens": [ 512, 512, 512 ], "actualGenerationTokens": [ 128, 128, 128 ], "failureDetail": "", "statistics": { "promptTokensPerSecond": { "min": 1965.8283739681322, "median": 1973.314011300306, "max": 1987.8477277580416 }, "generationTokensPerSecond": { "min": 47.707450401155796, "median": 47.70922467250623, "max": 47.7235311955947 }, "timeToFirstTokenSeconds": { "min": 0.2648635692894459, "median": 0.2665485143661499, "max": 0.2676740474998951 } } }, "evaluation": null, "timestamp": "2026-08-19T02:13:31.455226Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-124928.json", "serverLog": "logs/server-boundary-124928.log", "memoryLog": "logs/memory-boundary-124928.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-124928.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:c0be66a96c0138083910e87bc42ecc13e1fe3a1985212158298e5af66515ffeb" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-125952-1787105636", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "125952", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 125952, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 1, "measuredRuns": 0, "peakMemoryBytes": 25230835712, "peakMemoryMiB": 24062, "loadSeconds": 6.131869122385979, "promptTokensPerSecond": 0, "generationTokensPerSecond": 0, "timeToFirstTokenSeconds": 0, "outcome": "oom", "actualPromptTokens": [], "actualGenerationTokens": [], "failureDetail": "URLError(ConnectionRefusedError(111, 'Connection refused'))", "statistics": { "promptTokensPerSecond": null, "generationTokensPerSecond": null, "timeToFirstTokenSeconds": null } }, "evaluation": null, "timestamp": "2026-08-19T02:13:48.854566Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-125952.json", "serverLog": "logs/server-boundary-125952.log", "memoryLog": "logs/memory-boundary-125952.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-125952.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:d3f7e6f9d3fee41ec208865c344aec732f88720574a84c00ea0ee7179b07c375" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-126976-1787105611", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "126976", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 126976, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "boundary" }, "measurements": { "warmupRuns": 0, "measuredRuns": 0, "peakMemoryBytes": 25237127168, "peakMemoryMiB": 24068, "loadSeconds": 6.34332362934947, "promptTokensPerSecond": 0, "generationTokensPerSecond": 0, "timeToFirstTokenSeconds": 0, "outcome": "oom", "actualPromptTokens": [], "actualGenerationTokens": [], "failureDetail": "server exited with code -6", "statistics": { "promptTokensPerSecond": null, "generationTokensPerSecond": null, "timeToFirstTokenSeconds": null } }, "evaluation": null, "timestamp": "2026-08-19T02:13:23.607547Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-boundary-126976.json", "serverLog": "logs/server-boundary-126976.log", "memoryLog": "logs/memory-boundary-126976.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/boundary-126976.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:5807e24b5909abcb4486ca6a1e249c7ae955a9ac00ff71f01fc2e33aca91b6d9" }, { "schemaVersion": "1.0.0", "runId": "qwen38-27b-q4km-rtx4090-f16kv-131072-1787105551", "model": { "id": "qwen3.8-27b", "repository": "Qwen/Qwen3.8-27B", "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" }, "artifact": { "filename": "Qwen3.8-27B-Q4_K_M.gguf", "repository": "unsloth/Qwen3.8-27B-GGUF", "revision": "f1bfb127c64f7072bdd2cad55f258b9c8b2910fe", "sha256": "7e78da5d7e3ae28d178121f58646953305f3e5bd3cb46f4a75584e8b6c6fe169", "sizeBytes": 17106775008, "quantization": "q4-k-m", "contentHashSource": "Hugging Face x-linked-etag and independent local SHA-256" }, "runtime": { "name": "llama.cpp", "version": "version: 0.1.2-dev (build 1, commit 6d0549831)", "commit": "6d05498314db1b57f81c271080018aa2d0b89be9", "command": [ "/kingy-local-lab/llama.cpp/build/bin/llama-server", "-m", "/kingy-local-lab/models/Qwen3.8-27B-Q4_K_M.gguf", "-c", "131072", "-ngl", "999", "-ctk", "f16", "-ctv", "f16", "-b", "512", "-ub", "512", "--parallel", "1", "--host", "127.0.0.1", "--port", "8080", "--metrics", "--no-mmap" ] }, "environment": { "os": "Linux-6.14.0-29-generic-x86_64-with-glibc2.39", "hardware": "NVIDIA GeForce RTX 4090", "hardwareId": "rtx-4090-24gb", "gpuUuid": "GPU-26f05cd8-3563-5afe-e659-1dcadfa30718", "driverVersion": "580.65.06", "memoryBytes": 134217728000, "environmentLog": "logs/environment.txt", "cloudProvider": "RunPod Secure Cloud", "cloudRegion": "EUR-IS-1", "podId": "dto9mj08pd9xy0", "allocatedVcpu": 64, "cpu": "AMD EPYC 7702 64-Core Processor", "hostVisibleCpuCount": 256, "hostVisibleMemoryBytes": 540790554624, "allocationSource": "RunPod console: 64 vCPU and 125 GB; host-visible values retained separately", "gpuMemoryMiB": 24564, "cudaDriverCapability": "13.0", "cudaToolkit": "12.8.93" }, "configuration": { "contextTokens": 131072, "kvCacheType": "f16", "batchSize": 512, "microBatchSize": 512, "parallelism": 1, "gpuLayersRequested": 999, "offload": "all model layers and KV cache on GPU; no CPU layer offload", "phase": "fixed" }, "measurements": { "warmupRuns": 0, "measuredRuns": 0, "peakMemoryBytes": 16496197632, "peakMemoryMiB": 15732, "loadSeconds": 6.129787795245647, "promptTokensPerSecond": 0, "generationTokensPerSecond": 0, "timeToFirstTokenSeconds": 0, "outcome": "oom", "actualPromptTokens": [], "actualGenerationTokens": [], "failureDetail": "server exited with code 1", "statistics": { "promptTokensPerSecond": null, "generationTokensPerSecond": null, "timeToFirstTokenSeconds": null } }, "evaluation": null, "timestamp": "2026-08-19T02:12:24.276159Z", "evidenceState": "observed", "reviewer": "Kingy AI Local Lab evidence review", "rawEvidence": { "requestLog": "logs/raw-fixed-131072.json", "serverLog": "logs/server-fixed-131072.log", "memoryLog": "logs/memory-fixed-131072.csv", "archiveUrl": "/data/evidence/kingy-qwen38-4090-evidence.tgz", "archiveSha256": "028781080144a73a611a6b729edf1af8e8fd4173bd047ec3a3dfc4bab86359b9", "receiptUrl": "/data/receipts/fixed-131072.receipt.json" }, "review": { "status": "passed", "reviewedAt": "2026-08-19T02:20:00Z", "checks": [ "artifact size and SHA-256 match the pinned Hugging Face revision response", "single NVIDIA GeForce RTX 4090 with 24,564 MiB reported by nvidia-smi", "llama.cpp commit, driver, CUDA, OS, CPU and memory evidence retained", "FP16 K/V cache, one slot and -ngl 999; no CPU layer offload fallback", "one warm-up and three measured repetitions for every successful context", "512 prompt tokens and 128 generated tokens in every measured success", "CUDA OOM failures and the 1,024-token boundary bracket retained" ], "limitations": [ "RunPod host allocation is not a retail desktop platform and may differ in CPU, thermals and power behavior", "llama.cpp ignored the artifact's extra blk.64/MTP tensors; quality was not evaluated", "the measured boundary is an allocated-context VRAM-fit boundary; successful requests used 512 prompt tokens and 128 generated tokens, not a near-full-context prompt", "the boundary is established for this exact artifact, runtime commit, FP16 KV cache and batch settings only", "the upstream base revision is a pinned model record; the GGUF conversion lineage to that exact base commit was not independently established" ] }, "receiptHash": "sha256:fe18a36fb0689d1ffa2822fac09ec26d0e09ad2c2e773782d31880caff019e32" } ], "sources": [ { "id": "qwen38-hf-api", "publisher": "Hugging Face / Qwen", "title": "Qwen/Qwen3.8-27B registry record", "url": "https://huggingface.co/api/models/Qwen/Qwen3.8-27B", "appliesTo": [ "repository", "revision", "parameter count", "license", "last modified" ] }, { "id": "qwen38-config", "publisher": "Qwen", "title": "Qwen3.8-27B immutable config", "url": "https://huggingface.co/Qwen/Qwen3.8-27B/blob/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0/config.json", "appliesTo": [ "architecture", "layers", "attention layout", "KV heads", "head dimension", "native context" ] }, { "id": "qwen38-unsloth-q4km", "publisher": "Unsloth / Hugging Face", "title": "Qwen3.8-27B Q4_K_M GGUF at pinned revision", "url": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/blob/f1bfb127c64f7072bdd2cad55f258b9c8b2910fe/Qwen3.8-27B-Q4_K_M.gguf", "appliesTo": [ "artifact filename", "repository revision", "byte size", "content hash" ] }, { "id": "llamacpp-6d054983", "publisher": "ggml-org", "title": "llama.cpp benchmark runtime commit", "url": "https://github.com/ggml-org/llama.cpp/commit/6d05498314db1b57f81c271080018aa2d0b89be9", "appliesTo": [ "runtime source", "build revision", "command semantics" ] }, { "id": "qwen36-model-card", "publisher": "Qwen", "title": "Qwen3.6-27B model card", "url": "https://huggingface.co/Qwen/Qwen3.6-27B", "appliesTo": [ "parameter count", "architecture", "native context" ] }, { "id": "qwen3-8b-model-card", "publisher": "Qwen", "title": "Qwen3-8B model card", "url": "https://huggingface.co/Qwen/Qwen3-8B", "appliesTo": [ "parameter count", "layers", "attention heads", "native context" ] }, { "id": "nvidia-rtx4090", "publisher": "NVIDIA", "title": "GeForce RTX 4090 specifications", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/", "appliesTo": [ "memory capacity", "memory type", "architecture" ] }, { "id": "nvidia-rtx5090", "publisher": "NVIDIA", "title": "GeForce RTX 5090 specifications", "url": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/", "appliesTo": [ "memory capacity", "memory type", "architecture" ] }, { "id": "apple-mac-studio", "publisher": "Apple", "title": "Mac Studio technical specifications", "url": "https://www.apple.com/mac-studio/specs/", "appliesTo": [ "unified memory options", "memory bandwidth", "chip configurations" ] } ], "provenance": { "retrievedAt": "2026-08-19T02:20:00Z", "license": "CC-BY-4.0", "methodology": "/methodology" } }