v75 (S24/SM8650, soc 57): Qwen3.5-4B sharded-W8 decode bundle — Plane A greedy-correct, 3.8 tok/s

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 v81/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/qwen3.5-4b-1024.json

@@ -0,0 +1,70 @@
+{
+  "schema_version": 1,
+  "_comment": "Qwen3.5-4B GatedDeltaNet hybrid TEXT decoder for v81. W8 decode (the ~7GB fp16 decode exceeds device RAM; W8 ~3.5GB fits + ~2x faster) + fp16 lm-head. Recurrent gated delta-rule linear attn + gated full-attn; decode-over-prompt; qwen3_5_generate host-op.",
+  "model": {
+    "name": "qwen3.5-4b-1024",
+    "family": "llm",
+    "dsp_arch": "v75",
+    "tokenizer_pre": "qwen2"
+  },
+  "params": {
+    "hidden": 2560,
+    "vocab": 248320,
+    "n_layers": 32,
+    "max_ctx": 1024,
+    "kv_dim": 1024,
+    "head_dim": 256,
+    "rope_theta": 10000000.0,
+    "eos_token_id": 248044,
+    "gated_delta": {
+      "q_heads": 16,
+      "kv_heads": 4,
+      "lin_k_heads": 16,
+      "lin_v_heads": 32,
+      "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,
+        27,
+        31
+      ]
+    }
+  },
+  "artifacts": {
+    "contexts": {
+      "decode_s0": {
+        "bin": "qwen354b_decode_s0_w8.bin"
+      },
+      "decode_s1": {
+        "bin": "qwen354b_decode_s1_w8.bin"
+      },
+      "lmhead": {
+        "bin": "qwen354b_lmhead_f16.bin"
+      }
+    },
+    "embed": "qwen354b_embed_f16.bin",
+    "tokenizer": "tokenizer.json"
+  },
+  "plan": {
+    "steps": [
+      {
+        "host": "qwen3_5_generate",
+        "params": {
+          "decode": "qwen354b_decode_s0_w8,qwen354b_decode_s1_w8",
+          "lmhead": "qwen354b_lmhead_f16",
+          "max_new": 64,
+          "neg": -50000
+        }
+      }
+    ]
+  }
+}

v75/qwen354b_decode_s0_w8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0866ab249efd39cc44a05bffeb4e46eb7b814403f6d6fd2976eef7c27a78db0
+size 1804955648

v75/qwen354b_decode_s1_w8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f57603c60fdb78e0ea341179e01bab89e5676d69f01847d7b90323c6cc7d02eb
+size 1804959744

v75/qwen354b_embed_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c78988d979f52a340f39cd69edd8e4b0a80218a5fa8d95b6158bfce4a3de2a34
+size 1271398400

v75/qwen354b_lmhead_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b82afa1d43bcbaa99d429807e9f05c2d97e26baa8a4f41d878276f9ab44fb43
+size 1272709120

v75/tokenizer.json

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