Add generic chat block to v79 manifest (template-as-data) (#4)

- Add generic chat block to v79 manifest (template-as-data) (76bd75a82f6cd9b68791fbb8837b941a7a0594e3)


Co-authored-by: Aman <Aman0runanywhere@users.noreply.huggingface.co>

v79/qwen3.5-0.8b-1024.json

@@ -1,65 +1,79 @@
 {
-  "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 (~0.5GB, single context, under the 2GB model-lib limit) + fp16 lm-head, under a tolerant greedy_tol gate (W8 weights only \u2014 the fp32 SSM scan + host conv/SSM/KV state are untouched). Vision tower + MTP not exported (text-LLM path).",
-  "model": {
-    "name": "qwen3.5-0.8b-1024",
-    "family": "llm",
-    "dsp_arch": "v79",
-    "tokenizer_pre": "qwen2"
+ "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 (~0.5GB, single context, under the 2GB model-lib limit) + fp16 lm-head, under a tolerant greedy_tol gate (W8 weights only — the fp32 SSM scan + host conv/SSM/KV state are untouched). Vision tower + MTP not exported (text-LLM path).",
+ "model": {
+  "name": "qwen3.5-0.8b-1024",
+  "family": "llm",
+  "dsp_arch": "v79",
+  "tokenizer_pre": "qwen2"
+ },
+ "chat": {
+  "system": {
+   "prefix": "<|im_start|>system\n",
+   "suffix": "<|im_end|>\n"
   },
-  "params": {
-    "hidden": 1024,
-    "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": "qwen3508b_decode_w8.bin"
-      },
-      "lmhead": {
-        "bin": "qwen3508b_lmhead_f16.bin"
-      }
-    },
-    "embed": "qwen3508b_embed_f16.bin",
-    "tokenizer": "tokenizer.json"
+  "user": {
+   "prefix": "<|im_start|>user\n",
+   "suffix": "<|im_end|>\n"
   },
-  "plan": {
-    "steps": [
-      {
-        "host": "qwen3_5_generate",
-        "params": {
-          "decode": "qwen3508b_decode_w8",
-          "lmhead": "qwen3508b_lmhead_f16",
-          "max_new": 64,
-          "neg": -50000
-        }
-      }
-    ]
+  "assistant": {
+   "prefix": "<|im_start|>assistant\n",
+   "suffix": "<|im_end|>\n"
+  }
+ },
+ "params": {
+  "hidden": 1024,
+  "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": "qwen3508b_decode_w8.bin"
+   },
+   "lmhead": {
+    "bin": "qwen3508b_lmhead_f16.bin"
+   }
+  },
+  "embed": "qwen3508b_embed_f16.bin",
+  "tokenizer": "tokenizer.json"
+ },
+ "plan": {
+  "steps": [
+   {
+    "host": "qwen3_5_generate",
+    "params": {
+     "decode": "qwen3508b_decode_w8",
+     "lmhead": "qwen3508b_lmhead_f16",
+     "max_new": 64,
+     "neg": -50000
+    }
+   }
+  ]
+ }
+}