Add v75 (SM8650) bundle: decode-only MoE, device-validated (Paris/4 correct)

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 v79/experts_scale.f32 filter=lfs diff=lfs merge=lfs -text
 v79/embed_table.f32 filter=lfs diff=lfs merge=lfs -text
 v81/experts_scale.f32 filter=lfs diff=lfs merge=lfs -text
+v75/experts_scale.f32 filter=lfs diff=lfs merge=lfs -text

v75/a_hi.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64b53896cd67c3bb554f28238627fa865d261fbf0cf50de9e0707a1784775242
+size 680792064

v75/a_lo.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02cddcbbf710676bfa947ddcd0f0fddd370188e6e962729a90588486d98b4eac
+size 680792064

v75/embed_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:847f7968052e7f1182010b71e4cb015475441099207565979a9414c932dae3c0
+size 262668288

v75/experts_i8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f17f0ed5f7e564c7be8b175d0c885bc7ce3a4532117463ca0e9e064421bfad2
+size 2818572288

v75/experts_scale.f32

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e640a0681bec093b2a00ba2f91b44204c42598f7bdd13353cb516d2a11ccce8
+size 10223616

v75/ffn2.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:952a707d6d458cf9425283024c3ca38b8a23b673f01bf8e027af9fa450a5a74e
+size 110592

v75/lmhead_ar1.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f341c1b1663c5353126239dfeec734780c8aad48797cdda494895ac3976c09c
+size 263249920

v75/phimoe.json

@@ -0,0 +1,58 @@
+{
+  "schema_version": 1,
+  "_comment": "Phi-tiny-MoE on Hexagon v81 (SM8850 / soc_model 87) \u2014 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": "v75",
+    "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
+        }
+      }
+    ]
+  }
+}