GLM 5.3 Flash / DFlash2 inference runtime 0.5.15
This is a Docker-save runtime archive for two NVIDIA Thor T5000 nodes with driver 595.78 and SM 11.0. It contains no model weights. Download the target and DFlash2 snapshots separately from the manateelazycat model repositories. The corresponding LPK uses the selected Hugging Face / ModelScope source for both model files and this archive. Public downloads require no hub login.
The historical 0.5.0 reference environment was TensorFold 0.6.1 plus Mia 1.3.2 with the validated Thor decoder and communication patches. Default capacity is 1,048,576 tokens per request, a 2,684,928-token shared KV pool, and four streams. Reference acceptance includes empty-cache CUDA JIT, vision, structured output, tools, exact serial/drafted/concurrent token equality, restart, three-run benchmarks, and full 1M three-position retrieval with format-only guided regex. Version 0.5.1 repeats the local two-rank draft timing calibration in three independent batches and selects one complete measured table. It does not embed timing constants from the reference machines. Version 0.5.4 bundles the Thor Q4 prompt tiling and a four-rail TCP transport for 1–16 MiB prompt exchanges. The native library and launch integration are inside the image; no experiment directory or source bind mount is required. Both ranks discover their existing optical IPv4 addresses and match four distinct subnets through the control store. Incomplete four-link topology retains NCCL; decode and smaller/larger exchanges retain the existing path. Transport faults are reported as errors, never silently replaced with stale results. The sparse-attention experiment is not part of the released profile. Version 0.5.9 ports the Mia v1.10 recipe's 0085 sparse-selection and 0086 prompt scoring loops, shared-prefix copying and kept-state eviction/compaction fixes. It enables cross-rank nucleus candidate sampling and disables lone-stream cache relocation so an intervening conversation retains the long conversation's state. The frozen TensorFold 0.6.1 / Mia 1.3.2 runtime and Thor native libraries remain. Reasoning-effort tail relocation is implemented but off by default; cache-share growth is off, preserving the existing pool and eight kept states. The exact ported commit and patch list are included in runtime-manifest.json's profile. Version 0.5.10 rebases both the CLI and the active GLM CUDA backend to official TensorFold 0.6.6, retaining the v1.10 recipe and audited Thor native libraries. All adaptations are bundled inside the image; runtime source overlays are not required. See the pinned profile for revisions and the acceptance evidence. Version 0.5.14 bundles the validated 0101/0105/0106/0107 index port and its three precompiled SM110 modules. Thor kernels, four optical rails, model quantization, the 1M context budget and four streams retain their accepted configuration. The onepass attention dependency is compiled; its 17-24 local-head route is inactive for this TP2/H32 deployment. See profile.indexOptimizations for hashes. Version 0.5.15 includes the remaining applicable upstream recipe patches for the Thor TP2 configuration: 24 additional ports, 99 of 107 patches covered, and eight hardware-specific exclusions. It adds the CUDA SSD KV tier with 64 GiB per rank, persistent snapshots, clean-exit flushing and compatibility checks, plus sliced prefill and eight-stream backend capability. The default remains four streams. SM110 retains its measured prompt tile choices. The LPK archive packages the exact source-audited runtime already tested on the two target machines; this packaging step changes no inference code or binary. See profile.thorPatchIntegration and the linked acceptance evidence for the fixed upstream revision, adaptations and measured performance. The regex constrains labels and hex syntax, never the random expected values. An earlier 0.5.0 unguided 1M run retrieved all three values but omitted the labels, so its strict output-format check failed and its receipt is retained. See runtime-manifest.json for exact identities and measured performance; capacity is not a speed claim.
Check SHA256SUMS before docker load. Model volumes are mounted read-only; compiler caches are separate. Framework/code notices are preserved inside the archive. Component licenses apply independently; this archive grants no additional rights to model weights. DFlash2 weights are CC BY-NC-ND 4.0 and must be used without modification in the authorized noncommercial deployment.
Incremental archive installation
The required import order is: glm-5.3-flash-dual-t5000-runtime-0.5.14.tar.gz -> glm-5.3-flash-dual-t5000-runtime-0.5.15-source.tar.gz -> glm-5.3-flash-dual-t5000-runtime-0.5.15-native.tar.gz -> glm-5.3-flash-dual-t5000-runtime-0.5.15-incremental.tar.gz.
The LPK checks and imports missing dependencies from the deepest
base first, then imports the final incremental archive. Every
archive follows the selected Hugging Face / ModelScope source.
For manual installation, recursively follow base in
runtime-manifest.json. Download each archive at its exact
sources revision, verify its SHA256, and run docker load
in the order above.
Loading only the incremental archive into an empty Docker store
leaves required layers missing. The resulting complete image has
the exact accepted image ID, rootfs layers, code and native libraries.
Isolated empty-store import and every runtime file SHA were tested.