Upload v79/phimoe.json with huggingface_hub

v79/phimoe.json

@@ -1,9 +1,9 @@
 {
   "schema_version": 1,
-  "_comment": "Phi-tiny-MoE (microsoft/Phi-tiny-MoE-instruct) — FIRST MoE on v79, as a native QHexRT LLM via the phimoe_generate host-op. AR=1 KV-decode: a0..a31 are per-layer attn+router graphs (each holds its layer's KV cache) + host sparsemixer top-2 + ONE dynamic-weight expert FFN (ffndyn; the 2 router-selected experts are dequantized from host-RAM int8 each layer) + in-graph final LayerNorm + lm-head (lmhead_ar1). Quant = weight-only int8 experts + fp16 attn/router/lm-head (int16 acts break attention; W4 too coarse). 4 QNN contexts -> fits the HTP ~8-concurrent-context limit. Prompt: text (Phi-3 chat template [<|user|>] + spm(text) + [<|end|>,<|assistant|>]) OR a raw comma-separated id list. 100% greedy parity vs the fp32 oracle. MAXCTX 64 (the AR=1 KV width).",
+  "_comment": "Phi-tiny-MoE (microsoft/Phi-tiny-MoE-instruct) — FIRST MoE on v79, as a native QHexRT LLM via the phimoe_generate host-op. AR=1 KV-decode: a0..a31 are per-layer attn+router graphs (each holds its layer's KV cache) + host sparsemixer top-2 + ONE dynamic-weight expert FFN (ffndyn; the 2 router-selected experts are dequantized from host-RAM int8 each layer) + in-graph final LayerNorm + lm-head (lmhead_ar1). Quant = weight-only int8 experts + fp16 attn/router/lm-head (int16 acts break attention; W4 too coarse). 4 QNN contexts -> fits the HTP ~8-concurrent-context limit. Prompt: text (Phi-3 chat template [<|user|>] + spm(text) + [<|end|>,<|assistant|>]) OR a raw comma-separated id list. MAXCTX 2048 (16 Q-heads -> materialized+native-softmax wide attention is HTP-correct >512; verified 12/12 vs fp32 on a 529-token needle-recall prompt). Static MAXCTX-wide attention every token, so decode (~2 tok/s) is slower than a smaller window; prefill is decode-over-prompt (~0.45 s/prompt-token at 2048).",
   "model": { "name": "phi-tiny-moe", "family": "llm", "dsp_arch": "v79", "tokenizer_pre": "default" },
   "params": {
-    "hidden": 4096, "vocab": 32064, "n_layers": 32, "max_ctx": 64,
+    "hidden": 4096, "vocab": 32064, "n_layers": 32, "max_ctx": 2048,
     "kv_dim": 512, "head_dim": 128, "rope_theta": 10000, "eos_token_id": 32000
   },
   "artifacts": {