Skip to main content

Module native_asset

Module native_asset 

Source
Expand description

Native Qwen4Exp asset preparation.

The v4 P64 format intentionally uses 32-bit relative offsets, so a single P64 can never represent the 90+ GiB Qwen3.8 Flash Next checkpoint. The correct native artefact today is a compact .hmc contract which attests to the external GGUF and describes the PLE table’s direct NVMe byte span. Future sub-4-GiB P64 trunk shards, .q42 policy volumes, and .10d schedules are added as intact HMC entries without ever copying the PLE.

Structs§

ExternalGgufAsset
A verified external source. The HMC intentionally stores the identity and relative file name, not a host-specific absolute path.
Qwen4ExpNativeDescriptor
Cold activation data placed in the .hmc descriptor entry.
Qwen4ExpPleSpan
Directly addressable PLE tensor span inside the external GGUF payload.
Qwen4ExpPleStorage
The extracted PLE payload. It is the only large native artifact: a raw, row-addressable 26.8 GiB table on the chosen NVMe volume, never a second GGUF or a RAM/VRAM-resident buffer.

Enums§

Qwen4ExpNativeAssetError

Constants§

QWEN4EXP_NATIVE_FORMAT

Functions§

load_qwen4exp_native_hmc
Read the compact descriptor from a transparent HMC without loading any external model bytes. The caller resolves and validates the external GGUF with Qwen4ExpNativeDescriptor::verify_source_metadata or Qwen4ExpNativeDescriptor::verify_source_sha256.
prepare_qwen4exp_native_hmc
Create a compact, immutable .hmc activation package on an artifact volume while its GGUF trunk remains in place. Only the PLE span is extracted to ple_output; all other 60+ GiB stay in the source GGUF. An already-present PLE payload of the exact declared length is adopted without copying it again, allowing a compact HMC schema refresh without consuming disk space.