Hugging Face's logo

optimum-internal-testing
/
neuron-testing-cache

Model card Files
xet
 Community
optimum-internal-testing-user commited on
Commit
82c1a11
·
verified ·
1 Parent(s): cf8aece

Synchronizing local compiler cache.

Browse files
Files changed (6) hide show
  1. neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/c377610e3898d006f0a0.json +59 -0
  2. neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/granite/ibm-granite/granite-3.1-2b-instruct/c377610e3898d006f0a0.json +59 -0
  3. neuronxcc-2.21.33363.0+82129205/MODULE_4ff804c2533da46e613e+24129607/model.hlo_module.pb +1 -1
  4. neuronxcc-2.21.33363.0+82129205/MODULE_4ff804c2533da46e613e+24129607/model.neff +1 -1
  5. neuronxcc-2.21.33363.0+82129205/MODULE_5742ef8f5910f4e17377+a02c3a36/model.neff +1 -1
  6. neuronxcc-2.21.33363.0+82129205/MODULE_5742ef8f5910f4e17377+a02c3a36/wrapped_neff.hlo +1 -1
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/c377610e3898d006f0a0.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "ibm-granite/granite-3.1-2b-instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "GraniteForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.1,
10
+ "attention_multiplier": 0.015625,
11
+ "dtype": "bfloat16",
12
+ "embedding_multiplier": 12.0,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 2048,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 8192,
17
+ "logits_scaling": 8.0,
18
+ "max_position_embeddings": 131072,
19
+ "mlp_bias": false,
20
+ "model_type": "granite",
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 4,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
26
+ "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
27
+ "continuous_batching": true,
28
+ "ep_degree": 1,
29
+ "fused_qkv": true,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "max_batch_size": 4,
33
+ "max_context_length": 1024,
34
+ "max_topk": 256,
35
+ "n_active_tokens": 1024,
36
+ "neuronxcc_version": "2.21.33363.0+82129205",
37
+ "on_device_sampling": true,
38
+ "optimum_neuron_version": "0.4.6.dev4",
39
+ "output_logits": false,
40
+ "pp_degree": 1,
41
+ "prefill_chunk_size": 0,
42
+ "sequence_length": 1024,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": "trn1",
46
+ "torch_dtype": "bfloat16",
47
+ "tp_degree": 2
48
+ },
49
+ "num_attention_heads": 32,
50
+ "num_hidden_layers": 40,
51
+ "num_key_value_heads": 8,
52
+ "residual_multiplier": 0.22,
53
+ "rms_norm_eps": 1e-05,
54
+ "rope_scaling": null,
55
+ "rope_theta": 5000000.0,
56
+ "tie_word_embeddings": true,
57
+ "use_cache": true,
58
+ "vocab_size": 49155
59
+ }
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.6.dev4/granite/ibm-granite/granite-3.1-2b-instruct/c377610e3898d006f0a0.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_entry_class": "SingleModelCacheEntry",
3
+ "_model_id": "ibm-granite/granite-3.1-2b-instruct",
4
+ "_task": "text-generation",
5
+ "architectures": [
6
+ "GraniteForCausalLM"
7
+ ],
8
+ "attention_bias": false,
9
+ "attention_dropout": 0.1,
10
+ "attention_multiplier": 0.015625,
11
+ "dtype": "bfloat16",
12
+ "embedding_multiplier": 12.0,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 2048,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 8192,
17
+ "logits_scaling": 8.0,
18
+ "max_position_embeddings": 131072,
19
+ "mlp_bias": false,
20
+ "model_type": "granite",
21
+ "neuron": {
22
+ "_serialized_key": "NxDNeuronConfig",
23
+ "batch_size": 4,
24
+ "capacity_factor": null,
25
+ "checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
26
+ "checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
27
+ "continuous_batching": true,
28
+ "ep_degree": 1,
29
+ "fused_qkv": true,
30
+ "glu_mlp": true,
31
+ "local_ranks_size": 2,
32
+ "max_batch_size": 4,
33
+ "max_context_length": 1024,
34
+ "max_topk": 256,
35
+ "n_active_tokens": 1024,
36
+ "neuronxcc_version": "2.21.33363.0+82129205",
37
+ "on_device_sampling": true,
38
+ "optimum_neuron_version": "0.4.6.dev4",
39
+ "output_logits": false,
40
+ "pp_degree": 1,
41
+ "prefill_chunk_size": 0,
42
+ "sequence_length": 1024,
43
+ "speculation_length": 0,
44
+ "start_rank_id": 0,
45
+ "target": "trn1",
46
+ "torch_dtype": "bfloat16",
47
+ "tp_degree": 2
48
+ },
49
+ "num_attention_heads": 32,
50
+ "num_hidden_layers": 40,
51
+ "num_key_value_heads": 8,
52
+ "residual_multiplier": 0.22,
53
+ "rms_norm_eps": 1e-05,
54
+ "rope_scaling": null,
55
+ "rope_theta": 5000000.0,
56
+ "tie_word_embeddings": true,
57
+ "use_cache": true,
58
+ "vocab_size": 49155
59
+ }
neuronxcc-2.21.33363.0+82129205/MODULE_4ff804c2533da46e613e+24129607/model.hlo_module.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ba0dce0602506ec104796b8fd8725dee074d29c0e88ce5ce7f4b4f681250622b
3
  size 1050679
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:868cca689a34e99963ba4838928c3d05b3b7ff0c504770b6e3b057f912c499df
3
  size 1050679
neuronxcc-2.21.33363.0+82129205/MODULE_4ff804c2533da46e613e+24129607/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fb95dbd8cc3cab5b05fe9e7717acda2130bde6a0d8ac99b0a2fca725115f9ddf
3
  size 14050304
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2e25df46041906782db39a6b3e925eb663445c890658b31dd76e5b73d01f3f5b
3
  size 14050304
neuronxcc-2.21.33363.0+82129205/MODULE_5742ef8f5910f4e17377+a02c3a36/model.neff CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:12483685f9f33f8d93d63a0fd97933ff8dd168697144d649461cd3b5aca95c98
3
  size 5387264
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b7c5b6b3784fd9e10afff309616b6a27dd7b8810dc29cfe53645820d74ecbe9
3
  size 5387264
neuronxcc-2.21.33363.0+82129205/MODULE_5742ef8f5910f4e17377+a02c3a36/wrapped_neff.hlo CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:bd7c6b14f10695fed21a2043ca93c89feabc56c16c28f11ec89dc8b0c95cd73e
3
  size 5571457
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d4cd702720ca0453e414fb4c60a007f902afd4d799c17b4b0c3ef0249aae319
3
  size 5571457