GLM-4.6V-Flash-REFFT

Hexagon HTP serving bundles for ZhipuAI/GLM-4.6V-Flash.

Each artifact ending in .refft is one packaging variant for refft-hexagon (cli | serve).

Preferred layout (single-file RFTB)

GLM-4.6V-Flash-W4.refft   # one RFTB file (magic RFTB)

Field	Value
Source	GLM-4.6V-Flash W4DA16-PERF
Container	RFTB (general.scheme=hexagon.serving.v1)
Load	mmap BlobStore (no extract)
Runtime	generic Visa LM (host_ops.embed.placement=host_only)
Weights	multi-part paged (p768), resident_window=1
Task	causal_lm (text; vision not in this pack)

Device smoke (IQ-9075 / ubuntu-arm64-v73)

Skel: use interleaved partial-RoPE build (e.g. librefft_hexagon_v73_rope_safe_reint.so). soft_pair can run but may yield wrong first tokens.

# After the file is on device:
export REFFT_HEXAGON_MODULE_PATH=/librefft_hexagon_v73_rope.so
export REFFT_HEXAGON_ARCH_VERSION=73
export REFFT_HEXAGON_MAX_PERFORMANCE_MODE=1

refft-hexagon cli \
  --model /path/to/GLM-4.6V-Flash-W4.refft \
  --backend hexagon \
  --prompt "Who are you?" \
  --max_new_tokens 16

Verified reference: chat prompt “Who are you?” → first_token=151350 and Chinese <think>… prefix.

Docs

• docs/guides/refft_serving_bundle_rftb.md
• docs/guides/refft_generic_execution_contract.md
• docs/guides/refft_serving_pack_test_upload.md