Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files Community
tengomucho HF Staff commited on
Commit
89bba7b
·
verified ·
1 Parent(s): c62fabb

Synchronizing local compiler cache.

Browse files
Files changed (7) hide show
  1. neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev6/llama/unsloth/Llama-3.2-1B-Instruct/0652c1c39ac08c855ad8.json +78 -0
  2. neuronxcc-2.17.194.0+d312836f/MODULE_34a6b42796c8b4e2f58b+431f5505/model.neff +1 -1
  3. neuronxcc-2.17.194.0+d312836f/MODULE_b6cc4e3b2052c9ae8504+bfe5714b/model.hlo_module.pb +1 -1
  4. neuronxcc-2.17.194.0+d312836f/MODULE_b6cc4e3b2052c9ae8504+bfe5714b/model.neff +1 -1
  5. neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/model.hlo_module.pb +1 -1
  6. neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/model.neff +1 -1
  7. neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/wrapped_neff.hlo +1 -1
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev6/llama/unsloth/Llama-3.2-1B-Instruct/0652c1c39ac08c855ad8.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "unsloth/Llama-3.2-1B-Instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "LlamaForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.0,
10
+ "head_dim": 64,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 2048,
13
+ "initializer_range": 0.02,
14
+ "intermediate_size": 8192,
15
+ "max_position_embeddings": 131072,
16
+ "mlp_bias": false,
17
+ "model_type": "llama",
18
+ "neuron": {
19
+ "_serialized_key": "NxDNeuronConfig",
20
+ "async_mode": false,
21
+ "attn_kernel_enabled": false,
22
+ "batch_size": 4,
23
+ "capacity_factor": null,
24
+ "cc_pipeline_tiling_factor": 2,
25
+ "checkpoint_id": "unsloth/Llama-3.2-1B-Instruct",
26
+ "checkpoint_revision": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
27
+ "continuous_batching": true,
28
+ "enable_bucketing": false,
29
+ "ep_degree": 1,
30
+ "flash_decoding_enabled": false,
31
+ "fused_qkv": true,
32
+ "glu_mlp": true,
33
+ "is_chunked_prefill": false,
34
+ "local_ranks_size": 2,
35
+ "logical_nc_config": 1,
36
+ "max_batch_size": 4,
37
+ "max_context_length": 4096,
38
+ "max_topk": 256,
39
+ "mlp_kernel_enabled": false,
40
+ "mlp_kernel_fuse_residual_add": false,
41
+ "n_active_tokens": 4096,
42
+ "neuronxcc_version": "2.17.194.0+d312836f",
43
+ "num_cores_per_group": 1,
44
+ "on_device_sampling": false,
45
+ "optimum_neuron_version": "0.2.0.dev6",
46
+ "output_logits": false,
47
+ "padding_side": "right",
48
+ "pp_degree": 1,
49
+ "qk_layernorm": false,
50
+ "qkv_kernel_enabled": false,
51
+ "rpl_reduce_dtype": "float16",
52
+ "sequence_length": 4096,
53
+ "sequence_parallel_enabled": false,
54
+ "speculation_length": 0,
55
+ "start_rank_id": 0,
56
+ "target": null,
57
+ "torch_dtype": "float16",
58
+ "tp_degree": 2,
59
+ "vocab_parallel": false
60
+ },
61
+ "num_attention_heads": 32,
62
+ "num_hidden_layers": 16,
63
+ "num_key_value_heads": 8,
64
+ "pretraining_tp": 1,
65
+ "rms_norm_eps": 1e-05,
66
+ "rope_scaling": {
67
+ "factor": 32.0,
68
+ "high_freq_factor": 4.0,
69
+ "low_freq_factor": 1.0,
70
+ "original_max_position_embeddings": 8192,
71
+ "rope_type": "llama3"
72
+ },
73
+ "rope_theta": 500000.0,
74
+ "tie_word_embeddings": true,
75
+ "unsloth_fixed": true,
76
+ "use_cache": true,
77
+ "vocab_size": 128256
78
+ }
neuronxcc-2.17.194.0+d312836f/MODULE_34a6b42796c8b4e2f58b+431f5505/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:075eab47089f6b5408189c16efa63c02b2be7813b34cf32f7ac353459842c720
3
  size 1158144
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:277e40266261753d321701565433ece6105d089a926dc9b8d5205a103afa4c67
3
  size 1158144
neuronxcc-2.17.194.0+d312836f/MODULE_b6cc4e3b2052c9ae8504+bfe5714b/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4c31ca5c42669af732c52e7635faec4f9600d0cf781e381c4a2a7f791c10b3d4
3
  size 445394
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7ca1cbe7c0596bf195c92a6981d58f97a15f5ef962024a9c2af6284578036cd
3
  size 445394
neuronxcc-2.17.194.0+d312836f/MODULE_b6cc4e3b2052c9ae8504+bfe5714b/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b31e34c8894efaa104e01140f1968514edb44ca9a21a3bccafa8e11c5ce5b78d
3
  size 32072704
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ec00109ee082b942225b34bca892da58326723463e3dc5855748b4d2dbe95d6
3
  size 32072704
neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fd29b9fcad2a945198bae1058538f54be003576d1675a8d9da0832151db4194e
3
  size 378951
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc7169938d5d2448836a0ba2caee343313acd17800c1c8d5340847866063cca4
3
  size 378951
neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ca0984fe4b17c72150b3f1bb21b08b7e480ba250f6e1084471aa1f5f7a7e5084
3
  size 4404224
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b69e7f3bc5be603f95247402fad05e67dd40e922a65e527b24b3a99ced7904d
3
  size 4404224
neuronxcc-2.17.194.0+d312836f/MODULE_b94c33a1002b2de0e6f0+165e9558/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dc5f7c13ca6d344a2adce9d3a54e07eca1e4972c01f35ae91fc1d2f12b6d0380
3
  size 4473980
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fff35422bbcefb32702e89c4911dd8a1702ee1ee9f6d3d8121ded87253fb327
3
  size 4473980