Synchronizing local compiler cache.
Browse files- neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev2/inference/granite/ibm-granite/granite-3.1-2b-instruct/44dbd825dcd869d9d169.json +51 -0
- neuronxcc-2.17.194.0+d312836f/MODULE_20223e04b1bfd1a5fe7e+613edded/model.neff +1 -1
- neuronxcc-2.17.194.0+d312836f/MODULE_4f54951d7f1d19705651+613edded/model.neff +1 -1
neuronxcc-2.17.194.0+d312836f/0_REGISTRY/0.2.0.dev2/inference/granite/ibm-granite/granite-3.1-2b-instruct/44dbd825dcd869d9d169.json
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 |
+
"embedding_multiplier": 12.0,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 2048,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 8192,
|
| 16 |
+
"logits_scaling": 8.0,
|
| 17 |
+
"max_position_embeddings": 131072,
|
| 18 |
+
"mlp_bias": false,
|
| 19 |
+
"model_type": "granite",
|
| 20 |
+
"neuron": {
|
| 21 |
+
"all_reduce_dtype": null,
|
| 22 |
+
"allow_flash_attention": true,
|
| 23 |
+
"attention_layout": "HSB",
|
| 24 |
+
"attn_output_transposed": false,
|
| 25 |
+
"auto_cast_type": "bf16",
|
| 26 |
+
"backend": "hlo",
|
| 27 |
+
"batch_size": 4,
|
| 28 |
+
"checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
|
| 29 |
+
"checkpoint_revision": "374ef54e020a3ce208c65e96d6213922a87d8952",
|
| 30 |
+
"collectives_layout": "HSB",
|
| 31 |
+
"continuous_batching": true,
|
| 32 |
+
"fuse_qkv": true,
|
| 33 |
+
"group_query_attention": "shard-over-heads",
|
| 34 |
+
"log_softmax_scores": false,
|
| 35 |
+
"neuronxcc_version": "2.17.194.0+d312836f",
|
| 36 |
+
"optimum_neuron_version": "0.2.0.dev2",
|
| 37 |
+
"output_all_logits": false,
|
| 38 |
+
"sequence_length": 4096,
|
| 39 |
+
"tp_degree": 2
|
| 40 |
+
},
|
| 41 |
+
"num_attention_heads": 32,
|
| 42 |
+
"num_hidden_layers": 40,
|
| 43 |
+
"num_key_value_heads": 8,
|
| 44 |
+
"residual_multiplier": 0.22,
|
| 45 |
+
"rms_norm_eps": 1e-05,
|
| 46 |
+
"rope_scaling": null,
|
| 47 |
+
"rope_theta": 5000000.0,
|
| 48 |
+
"tie_word_embeddings": true,
|
| 49 |
+
"use_cache": true,
|
| 50 |
+
"vocab_size": 49155
|
| 51 |
+
}
|
neuronxcc-2.17.194.0+d312836f/MODULE_20223e04b1bfd1a5fe7e+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 3851264
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f2ef11d91a58bd2be2fbdb1358eae675ec9fa8061da399cfa18b1f1262a46038
|
| 3 |
size 3851264
|
neuronxcc-2.17.194.0+d312836f/MODULE_4f54951d7f1d19705651+613edded/model.neff
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 5377024
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ea67c6c697fc4e9d1d721510e7a35e0807142ab68095ecd0a1c6f0b3c47391af
|
| 3 |
size 5377024
|