DeepSeek V4.1 Flash TensorFold runtime 0.1.9
For LPK 0.1.15: two NVIDIA Thor nodes, TP2, four optical TCP rails, eight shared request slots. The default total input/output context is 1,048,576 tokens. The KV pool is shared, rather than eight physically allocated 1M contexts. Includes the qualified upstream stability ports and Thor prefill kernels. Inference processes disable anonymous transparent huge pages to avoid synchronous compaction under memory pressure; host-wide policy is unchanged. Performance uses the frozen 1M-configured workload with matched actual input lengths, output budgets, concurrency and cache policy. Every retained median throughput and TTFT P50/P95 gate meets the original 2% tolerance. The failed earlier immutable run remains in the project evidence. Actual 64K, 128K, 256K, 512K and 1M capacity plus three-position retrieval were checked separately from performance regression. Replay prefill remains approximate, matching the frozen reference. This Docker-save archive contains no model weights. Fetch the separately pinned target and original Engram assets using the LPK manifests; verified local TP2 preparation is supported. Native CUDA modules are precompiled. Minimal C/Triton tooling for cold caches and dependency licenses/notices are retained. Verify SHA256SUMS before docker load.