v75 (Galaxy S24 / SM8650, soc_model 57): Qwen3.5-2B W8-decode bundle — Plane A 6/6, 10.5 tok/s

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 v81/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 v81-w8/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 v79/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+v75/tokenizer.json filter=lfs diff=lfs merge=lfs -text

v75/qwen3.5-2b-1024.json

@@ -0,0 +1,65 @@
+{
+  "schema_version": 1,
+  "_comment": "Qwen3.5 (qwen3_5) GatedDeltaNet hybrid TEXT decoder for v81. Most layers are a recurrent gated delta-rule linear attention (+ short conv state), 1-in-4 is gated softmax attention (partial-rotary M-RoPE, output gate). Decode-over-prompt seed; W8 decode (~1.3x faster, ~1.4GB \u2014 fits a single context and avoids the fp16 decode's >2GB large-code-model build) + fp16 lm-head. Device-validated v81 single-prompt greedy-exact 64/64 at 14.4 tok/s, length-swept suite 8/10 greedy_tol (W8 argmax-drift on 2 prompts, output stays coherent) under a tolerant greedy_tol gate. Vision tower + MTP not exported (text-LLM path).",
+  "model": {
+    "name": "qwen3.5-2b-1024",
+    "family": "llm",
+    "dsp_arch": "v75",
+    "tokenizer_pre": "qwen2"
+  },
+  "params": {
+    "hidden": 2048,
+    "vocab": 248320,
+    "n_layers": 24,
+    "max_ctx": 1024,
+    "kv_dim": 512,
+    "head_dim": 256,
+    "rope_theta": 10000000.0,
+    "eos_token_id": 248044,
+    "gated_delta": {
+      "q_heads": 8,
+      "kv_heads": 2,
+      "lin_k_heads": 16,
+      "lin_v_heads": 16,
+      "lin_k_dim": 128,
+      "lin_v_dim": 128,
+      "conv_kernel": 4,
+      "full_every": 4,
+      "rot_dim": 64,
+      "attn_gate": true,
+      "full_attn_idx": [
+        3,
+        7,
+        11,
+        15,
+        19,
+        23
+      ]
+    }
+  },
+  "artifacts": {
+    "contexts": {
+      "decode": {
+        "bin": "qwen352b_decode_w8.bin"
+      },
+      "lmhead": {
+        "bin": "qwen352b_lmhead_f16.bin"
+      }
+    },
+    "embed": "qwen352b_embed_f16.bin",
+    "tokenizer": "tokenizer.json"
+  },
+  "plan": {
+    "steps": [
+      {
+        "host": "qwen3_5_generate",
+        "params": {
+          "decode": "qwen352b_decode_w8",
+          "lmhead": "qwen352b_lmhead_f16",
+          "max_new": 64,
+          "neg": -50000
+        }
+      }
+    ]
+  }
+}

v75/qwen352b_decode_w8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8d4efd3117830114ce1ecd1a9c692e1ea71f8faebffbd95e5a7cbd1f1a3d6ff
+size 1392832512

v75/qwen352b_embed_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1014821eb1ecd1554be53855305701a402daced03fbc108ef7105223cdcd5d36
+size 1017118720

v75/qwen352b_lmhead_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81113c6d9d7f1e810704dcf3ff1426b5de304391ce23234241ce213e4faa1e29
+size 1018429440

v75/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325