Synchronizing local compiler cache.
Browse files- neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.0.dev3/granite/ibm-granite/granite-3.1-2b-instruct/2ccaa328023740986e21.json +73 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_6eee1d2655983d3e332f+84f3e719/model.hlo_module.pb +1 -1
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_6eee1d2655983d3e332f+84f3e719/model.neff +1 -1
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_c58c3b1c9745cdc8b7c7+5be477de/model.neff +1 -1
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_c58c3b1c9745cdc8b7c7+5be477de/wrapped_neff.hlo +1 -1
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_d84f59502564753faaf1+431f5505/model.neff +1 -1
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.0.dev3/granite/ibm-granite/granite-3.1-2b-instruct/2ccaa328023740986e21.json
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
"_serialized_key": "NxDNeuronConfig",
|
22 |
+
"async_mode": false,
|
23 |
+
"attn_kernel_enabled": false,
|
24 |
+
"batch_size": 4,
|
25 |
+
"capacity_factor": null,
|
26 |
+
"cc_pipeline_tiling_factor": 2,
|
27 |
+
"checkpoint_id": "ibm-granite/granite-3.1-2b-instruct",
|
28 |
+
"checkpoint_revision": "bbc2aed595bd38bd770263dc3ab831db9794441d",
|
29 |
+
"continuous_batching": true,
|
30 |
+
"enable_bucketing": false,
|
31 |
+
"ep_degree": 1,
|
32 |
+
"flash_decoding_enabled": false,
|
33 |
+
"fused_qkv": true,
|
34 |
+
"glu_mlp": true,
|
35 |
+
"is_chunked_prefill": false,
|
36 |
+
"local_ranks_size": 2,
|
37 |
+
"logical_nc_config": 1,
|
38 |
+
"max_batch_size": 4,
|
39 |
+
"max_context_length": 4096,
|
40 |
+
"max_topk": 256,
|
41 |
+
"mlp_kernel_enabled": false,
|
42 |
+
"mlp_kernel_fuse_residual_add": false,
|
43 |
+
"n_active_tokens": 4096,
|
44 |
+
"neuronxcc_version": "2.19.8089.0+8ab9f450",
|
45 |
+
"num_cores_per_group": 1,
|
46 |
+
"on_device_sampling": false,
|
47 |
+
"optimum_neuron_version": "0.3.0.dev3",
|
48 |
+
"output_logits": false,
|
49 |
+
"padding_side": "right",
|
50 |
+
"pp_degree": 1,
|
51 |
+
"qk_layernorm": false,
|
52 |
+
"qkv_kernel_enabled": false,
|
53 |
+
"rpl_reduce_dtype": "bfloat16",
|
54 |
+
"sequence_length": 4096,
|
55 |
+
"sequence_parallel_enabled": false,
|
56 |
+
"speculation_length": 0,
|
57 |
+
"start_rank_id": 0,
|
58 |
+
"target": null,
|
59 |
+
"torch_dtype": "bfloat16",
|
60 |
+
"tp_degree": 2,
|
61 |
+
"vocab_parallel": false
|
62 |
+
},
|
63 |
+
"num_attention_heads": 32,
|
64 |
+
"num_hidden_layers": 40,
|
65 |
+
"num_key_value_heads": 8,
|
66 |
+
"residual_multiplier": 0.22,
|
67 |
+
"rms_norm_eps": 1e-05,
|
68 |
+
"rope_scaling": null,
|
69 |
+
"rope_theta": 5000000.0,
|
70 |
+
"tie_word_embeddings": true,
|
71 |
+
"use_cache": true,
|
72 |
+
"vocab_size": 49155
|
73 |
+
}
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_6eee1d2655983d3e332f+84f3e719/model.hlo_module.pb
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 1072456
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:becf61930183ae1f6829d15b40543f541644d4b2e429c6efb218294c3a76cc60
|
3 |
size 1072456
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_6eee1d2655983d3e332f+84f3e719/model.neff
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 6605824
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:012b7bb8b3bc341342a6e0ce448375211809ba04d8d946b35bad7fcfc481fe3c
|
3 |
size 6605824
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_c58c3b1c9745cdc8b7c7+5be477de/model.neff
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 4885504
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a3de535d47ed1e8fe08e460d8839b3e671091bbbaab3ecae73f1b07be69f90b8
|
3 |
size 4885504
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_c58c3b1c9745cdc8b7c7+5be477de/wrapped_neff.hlo
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 5058549
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:66ff8746b41a1519a57c2af87dfd36e81d9c0bef96a52cfb1e0a7256d50aac11
|
3 |
size 5058549
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_d84f59502564753faaf1+431f5505/model.neff
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 2274304
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5237cf9ea1f91eab21026e48355b4d344369c5c5910ec1854933531050656867
|
3 |
size 2274304
|