Skip to content

Model implementation

models/ holds end-to-end LLM kernels, one flat directory per model build. Files ending in _draft.py are work in progress and excluded from CI.

Directory What it implements pypto-serving
qwen3_14b Qwen3-14B BF16 prefill and decode with the serving contract, plus A8W8 and TurboQuant variants and the sampling components Supported — one-card A2/A3 accuracy job on relevant PRs
deepseek_v4_flash_mtp DeepSeek V4-Flash at MTP = 1, batch 4 per card: operators, layer and MTP compositions, prefill/decode full forwards Supported — eight-card accuracy job on relevant PRs
deepseek_v4_pro Ascend A5 DeepSeek V4-Pro with an optional Flash preset, quantized Hybrid MXFP8-MXFP4 Not supported
deepseek_v4_flash_dspark The V4-Flash operators re-sized to batch 64 per card and S = 8 DSpark speculation; under development, operators only Not supported
deepseek_v3_2 DeepSeek V3.2-EXP as a front/back split of one layer Not supported
qwen3_32b Qwen3-32B single-layer decode in two tensor layouts Not supported

The linked directories have a page covering their deployment configuration and how their files compose. The rest are kernel harnesses validated against the Golden Harness only.

Entry points take script-specific platform and device arguments; inspect --help, the platform guide, and the Golden Harness guide.

python models/qwen3_14b/decode_fwd.py -p a2a3 -d 0
python models/deepseek_v4_flash_mtp/decode_layer.py -p a2a3 --ep 2 -d 0,1