v75 (Galaxy S24 / SM8650, soc_model 57): Canary-Qwen-2.5B LLM path (LoRA-merged Qwen3-1.7B backbone) — runs 13.4 tok/s; NOT a standalone text LLM (needs audio)

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 v81/canary/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 v81/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+v75/tokenizer.json filter=lfs diff=lfs merge=lfs -text

v75/canary-qwen-2.5b-llm.json

@@ -0,0 +1,49 @@
+{
+  "schema_version": 1,
+  "_comment": "Canary-Qwen-2.5B decoder as a standalone Qwen3-class native QHexRT LLM. Same architecture as Qwen3-0.6B / InternVL3.5-1B's LLM backbone (16q/8kv, head_dim 128, QK-norm, standard 1D RoPE theta 1e6, tied lm-head) \u2014 the proven 16-head materialized/gqa attention that runs coherently at MAXCTX>=1024. W8 decode + NPU tied lm-head + host embed lookup, driven by qwen3_generate. Device-validated on v81 (SM8850 / soc_model 87); W8 device gate is greedy_tol. See recipes/canary-qwen/DEVICE.md.",
+  "model": {
+    "name": "canary-qwen-2.5b-llm",
+    "family": "llm",
+    "dsp_arch": "v75",
+    "tokenizer_pre": "qwen2"
+  },
+  "params": {
+    "hidden": 2048,
+    "vocab": 151936,
+    "n_layers": 28,
+    "max_ctx": 1024,
+    "kv_dim": 1024,
+    "head_dim": 128,
+    "rope_theta": 1000000.0,
+    "eos_token_id": 151645
+  },
+  "artifacts": {
+    "contexts": {
+      "prefill": {
+        "bin": "canaryqwen_pf_f16.bin"
+      },
+      "decode": {
+        "bin": "canaryqwen_decn1k_w8.bin"
+      },
+      "lmhead": {
+        "bin": "canaryqwen_lmh_w8.bin"
+      }
+    },
+    "embed": "canaryqwen_embed_f16.bin",
+    "tokenizer": "tokenizer.json"
+  },
+  "plan": {
+    "steps": [
+      {
+        "host": "qwen3_generate",
+        "params": {
+          "prefill": "canaryqwen_pf_f16",
+          "decode": "canaryqwen_decn1k_w8",
+          "lmhead": "canaryqwen_lmh_w8",
+          "max_new": 64,
+          "neg": -50000
+        }
+      }
+    ]
+  }
+}

v75/canaryqwen_decn1k_w8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:908ecd555e09b997223ef2fe3e138f3a3b284d600dfd07e4160ee2e1c0a97459
+size 1419825152

v75/canaryqwen_embed_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bccd2bc5ec321a16d2c273d27634c9fe272fa5814fecc1fa48790fe7793edead
+size 622329856

v75/canaryqwen_lmh_w8.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ce6e53b89503461c722d61a3bac26369c435685413035fb2da6d9514e63593e
+size 313798656

v75/canaryqwen_pf_f16.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80844a8dcffa16d59b7aabf3e4e9b8f4d32879565ca40d14e4c7fb700f7c3b78
+size 2837450752

v75/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33ec919e3bb44401e327be3cd3461253ef9bdcc866a8585e001e8a33de7c17eb
+size 11422847