sm280299 commited on
Commit
3b770d6
·
verified ·
1 Parent(s): 01d77ac

Scrub internal engineering notes from manifest _comment

Browse files
Files changed (1) hide show
  1. v81/phimoe.json +1 -1
v81/phimoe.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "schema_version": 1,
3
- "_comment": "Phi-tiny-MoE on Hexagon v81 (SM8850 / soc_model 87) — DECODE-ONLY variant. The full 7-context bundle (64 graphs) crashes the v81 cDSP unsigned-PD in Graph::setup_vtcm (KB phimoe-v81-cdsp-unsigned-pd-vtcm-crash). This drops the 3 batched-prefill contexts (pf_lo/pf_hi/ffn_pf); the phimoe_generate host-op auto-falls-back to decode-over-prompt when they are absent (use_pf=false). 34 graphs / 4 contexts (~1.6 GB) fits the unsigned PD. Cost: slower TTFT on long prompts (decode-over-prompt); fine for short prompts. Decode: per-layer GQA-native attn+router a0..a31 (a_lo/a_hi) + host sparsemixer top-2 + fused 2-expert FFN (ffn2) + in-graph lm-head (lmhead_ar1). int8 experts + fp16 attn/router/lm-head.",
4
  "model": { "name": "phi-tiny-moe", "family": "llm", "dsp_arch": "v81", "tokenizer_pre": "default" },
5
  "params": {
6
  "hidden": 4096, "vocab": 32064, "n_layers": 32, "max_ctx": 2048,
 
1
  {
2
  "schema_version": 1,
3
+ "_comment": "phi-tiny-moe - Qualcomm Hexagon v81 (QHexRT) deployment manifest.",
4
  "model": { "name": "phi-tiny-moe", "family": "llm", "dsp_arch": "v81", "tokenizer_pre": "default" },
5
  "params": {
6
  "hidden": 4096, "vocab": 32064, "n_layers": 32, "max_ctx": 2048,