Hugging Face's logo

runanywhere
/
qwen3_5_4b_HNPU

Image-Text-to-Text
hnpu
hexagon
npu
vlm
v79
v81
Model card Files
xet
 Community
3
Aman0runanywhere commited on
Commit
d5c4ba0
·
verified ·
1 Parent(s): bf2080a

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

Browse files
Files changed (1) hide show
  1. v79/qwen3.5-4b-1024.json +80 -66
v79/qwen3.5-4b-1024.json CHANGED
@@ -1,70 +1,84 @@
1
  {
2
- "schema_version": 1,
3
- "_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.",
4
- "model": {
5
- "name": "qwen3.5-4b-1024",
6
- "family": "llm",
7
- "dsp_arch": "v79",
8
- "tokenizer_pre": "qwen2"
 
 
 
 
 
9
  },
10
- "params": {
11
- "hidden": 2560,
12
- "vocab": 248320,
13
- "n_layers": 32,
14
- "max_ctx": 1024,
15
- "kv_dim": 1024,
16
- "head_dim": 256,
17
- "rope_theta": 10000000.0,
18
- "eos_token_id": 248044,
19
- "gated_delta": {
20
- "q_heads": 16,
21
- "kv_heads": 4,
22
- "lin_k_heads": 16,
23
- "lin_v_heads": 32,
24
- "lin_k_dim": 128,
25
- "lin_v_dim": 128,
26
- "conv_kernel": 4,
27
- "full_every": 4,
28
- "rot_dim": 64,
29
- "attn_gate": true,
30
- "full_attn_idx": [
31
- 3,
32
- 7,
33
- 11,
34
- 15,
35
- 19,
36
- 23,
37
- 27,
38
- 31
39
- ]
40
- }
41
- },
42
- "artifacts": {
43
- "contexts": {
44
- "decode_s0": {
45
- "bin": "qwen354b_decode_s0_w8.bin"
46
- },
47
- "decode_s1": {
48
- "bin": "qwen354b_decode_s1_w8.bin"
49
- },
50
- "lmhead": {
51
- "bin": "qwen354b_lmhead_f16.bin"
52
- }
53
- },
54
- "embed": "qwen354b_embed_f16.bin",
55
- "tokenizer": "tokenizer.json"
56
  },
57
- "plan": {
58
- "steps": [
59
- {
60
- "host": "qwen3_5_generate",
61
- "params": {
62
- "decode": "qwen354b_decode_s0_w8,qwen354b_decode_s1_w8",
63
- "lmhead": "qwen354b_lmhead_f16",
64
- "max_new": 64,
65
- "neg": -50000
66
- }
67
- }
68
- ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69
  }
70
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  {
2
+ "schema_version": 1,
3
+ "_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.",
4
+ "model": {
5
+ "name": "qwen3.5-4b-1024",
6
+ "family": "llm",
7
+ "dsp_arch": "v79",
8
+ "tokenizer_pre": "qwen2"
9
+ },
10
+ "chat": {
11
+ "system": {
12
+ "prefix": "<|im_start|>system\n",
13
+ "suffix": "<|im_end|>\n"
14
  },
15
+ "user": {
16
+ "prefix": "<|im_start|>user\n",
17
+ "suffix": "<|im_end|>\n"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18
  },
19
+ "assistant": {
20
+ "prefix": "<|im_start|>assistant\n",
21
+ "suffix": "<|im_end|>\n"
22
+ }
23
+ },
24
+ "params": {
25
+ "hidden": 2560,
26
+ "vocab": 248320,
27
+ "n_layers": 32,
28
+ "max_ctx": 1024,
29
+ "kv_dim": 1024,
30
+ "head_dim": 256,
31
+ "rope_theta": 10000000.0,
32
+ "eos_token_id": 248044,
33
+ "gated_delta": {
34
+ "q_heads": 16,
35
+ "kv_heads": 4,
36
+ "lin_k_heads": 16,
37
+ "lin_v_heads": 32,
38
+ "lin_k_dim": 128,
39
+ "lin_v_dim": 128,
40
+ "conv_kernel": 4,
41
+ "full_every": 4,
42
+ "rot_dim": 64,
43
+ "attn_gate": true,
44
+ "full_attn_idx": [
45
+ 3,
46
+ 7,
47
+ 11,
48
+ 15,
49
+ 19,
50
+ 23,
51
+ 27,
52
+ 31
53
+ ]
54
  }
55
+ },
56
+ "artifacts": {
57
+ "contexts": {
58
+ "decode_s0": {
59
+ "bin": "qwen354b_decode_s0_w8.bin"
60
+ },
61
+ "decode_s1": {
62
+ "bin": "qwen354b_decode_s1_w8.bin"
63
+ },
64
+ "lmhead": {
65
+ "bin": "qwen354b_lmhead_f16.bin"
66
+ }
67
+ },
68
+ "embed": "qwen354b_embed_f16.bin",
69
+ "tokenizer": "tokenizer.json"
70
+ },
71
+ "plan": {
72
+ "steps": [
73
+ {
74
+ "host": "qwen3_5_generate",
75
+ "params": {
76
+ "decode": "qwen354b_decode_s0_w8,qwen354b_decode_s1_w8",
77
+ "lmhead": "qwen354b_lmhead_f16",
78
+ "max_new": 64,
79
+ "neg": -50000
80
+ }
81
+ }
82
+ ]
83
+ }
84
+ }