Upload v81/phimoe.json with huggingface_hub

v81/phimoe.json

@@ -0,0 +1,28 @@
+{
+  "schema_version": 1,
+  "_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.",
+  "model": { "name": "phi-tiny-moe", "family": "llm", "dsp_arch": "v81", "tokenizer_pre": "default" },
+  "params": {
+    "hidden": 4096, "vocab": 32064, "n_layers": 32, "max_ctx": 2048,
+    "kv_dim": 512, "head_dim": 128, "rope_theta": 10000, "eos_token_id": 32000
+  },
+  "artifacts": {
+    "contexts": {
+      "a_lo":   { "bin": "a_lo.bin" },
+      "a_hi":   { "bin": "a_hi.bin" },
+      "ffn":    { "bin": "ffn2.bin" },
+      "lmhead": { "bin": "lmhead_ar1.bin" }
+    },
+    "embed": "embed_f16.bin",
+    "tokenizer": "tokenizer.json",
+    "fixture_dir": "."
+  },
+  "plan": {
+    "steps": [
+      { "host": "phimoe_generate",
+        "params": { "experts": "experts_i8.bin", "scales": "experts_scale.f32",
+                    "n_experts": 16, "ffn": 448, "jitter": 0.01, "neg": -50,
+                    "user_id": 32010, "end_id": 32007, "assistant_id": 32001, "max_new": 24 } }
+    ]
+  }
+}