Types & classes — github.com/0xClandestine/mirror-sd

CodeHub MCPcopy Create free account

hub / github.com/0xClandestine/mirror-sd / types & classes

Types & classes21 in github.com/0xClandestine/mirror-sd

⨍ Functions 277◇ Types & classes 21

↓ 6 callersClassANEDraftModel

mirror_sd/ane_model.py:46

↓ 4 callersClassSpecDecodeStats

mirror_sd/generate.py:239

↓ 2 callersClassDFlashKVCache

KV cache for the DFlash draft model. Unlike the standard KVCache, this supports: - Explicit position tracking (not offset-based) for correct

mirror_sd/dflash.py:139

↓ 1 callersClassDFlashConfig

mirror_sd/dflash.py:35

↓ 1 callersClassDFlashDraftModel

mirror_sd/dflash.py:365

↓ 1 callersClassDFlashTrainModel

DFlash model extended for training with block construction and loss.

mirror_sd/train.py:53

↓ 1 callersClassQwen3DFlashAttention

mirror_sd/dflash.py:236

↓ 1 callersClassQwen3DFlashDecoderLayer

mirror_sd/dflash.py:333

↓ 1 callersClassQwen3MLP

mirror_sd/dflash.py:322

↓ 1 callersClassSpecServer

mirror_sd/server.py:33

↓ 1 callersClassTrainConfig

mirror_sd/train.py:226

↓ 1 callersClass_Measurement

scripts/ane_meter.py:91

↓ 1 callersClass_ProfiledANEDraftModel

Wraps ANEDraftModel, timing every forward call at the phase and kernel level.

scripts/bench_ane_profile.py:41

ane/src/wrapper.rs:183

ane/src/wrapper.rs:16

ClassDFlashDims

ane/src/dflash.rs:23

mirror_sd/server.py:170

Extract all weight data while holding the GIL, then release it for compilation.

ane/src/wrapper.rs:452

ClassQ8LayerWeights

ane/src/wrapper.rs:369

ClassQ8LayerWeights

Per-layer quantized projection weights for a transformer block.

ane/src/dflash.rs:13

ane/src/dflash.rs:5