Synchronizing local compiler cache.
Browse files- .gitattributes +3 -0
- neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.5.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/8a4430107c8609f149e2.json +58 -0
- neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.5.dev2/granite/ibm-granite/granite-3.1-2b-instruct/8a4430107c8609f149e2.json +58 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/compile_flags.json +1 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.done +0 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.hlo_module.pb +3 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.neff +3 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/wrapped_neff.hlo +3 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/compile_flags.json +1 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.done +0 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.hlo_module.pb +3 -0
- neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.neff +3 -0
.gitattributes
CHANGED
|
@@ -6353,3 +6353,6 @@ neuronxcc-2.21.33363.0+82129205/MODULE_169cfc2b665743de4e2b+ac10809c/wrapped_nef
|
|
| 6353 |
neuronxcc-2.21.33363.0+82129205/MODULE_10bbf823df24f9a5d76f+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 6354 |
neuronxcc-2.21.33363.0+82129205/MODULE_944d1c424c0b2150bceb+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 6355 |
neuronxcc-2.21.33363.0+82129205/MODULE_944d1c424c0b2150bceb+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 6353 |
neuronxcc-2.21.33363.0+82129205/MODULE_10bbf823df24f9a5d76f+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 6354 |
neuronxcc-2.21.33363.0+82129205/MODULE_944d1c424c0b2150bceb+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 6355 |
neuronxcc-2.21.33363.0+82129205/MODULE_944d1c424c0b2150bceb+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 6356 |
+
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 6357 |
+
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 6358 |
+
neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text
|
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.5.dev2/0cc25526e2cfc37a8875a3752f33c4d7505d8a07b869d0f3f41915cf6e763b74/8a4430107c8609f149e2.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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": 4,
|
| 32 |
+
"max_batch_size": 4,
|
| 33 |
+
"max_context_length": 4096,
|
| 34 |
+
"max_topk": 256,
|
| 35 |
+
"n_active_tokens": 4096,
|
| 36 |
+
"neuronxcc_version": "2.21.33363.0+82129205",
|
| 37 |
+
"on_device_sampling": true,
|
| 38 |
+
"optimum_neuron_version": "0.4.5.dev2",
|
| 39 |
+
"output_logits": false,
|
| 40 |
+
"pp_degree": 1,
|
| 41 |
+
"sequence_length": 4096,
|
| 42 |
+
"speculation_length": 0,
|
| 43 |
+
"start_rank_id": 0,
|
| 44 |
+
"target": "trn2",
|
| 45 |
+
"torch_dtype": "bfloat16",
|
| 46 |
+
"tp_degree": 4
|
| 47 |
+
},
|
| 48 |
+
"num_attention_heads": 32,
|
| 49 |
+
"num_hidden_layers": 40,
|
| 50 |
+
"num_key_value_heads": 8,
|
| 51 |
+
"residual_multiplier": 0.22,
|
| 52 |
+
"rms_norm_eps": 1e-05,
|
| 53 |
+
"rope_scaling": null,
|
| 54 |
+
"rope_theta": 5000000.0,
|
| 55 |
+
"tie_word_embeddings": true,
|
| 56 |
+
"use_cache": true,
|
| 57 |
+
"vocab_size": 49155
|
| 58 |
+
}
|
neuronxcc-2.21.33363.0+82129205/0_REGISTRY/0.4.5.dev2/granite/ibm-granite/granite-3.1-2b-instruct/8a4430107c8609f149e2.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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": 4,
|
| 32 |
+
"max_batch_size": 4,
|
| 33 |
+
"max_context_length": 4096,
|
| 34 |
+
"max_topk": 256,
|
| 35 |
+
"n_active_tokens": 4096,
|
| 36 |
+
"neuronxcc_version": "2.21.33363.0+82129205",
|
| 37 |
+
"on_device_sampling": true,
|
| 38 |
+
"optimum_neuron_version": "0.4.5.dev2",
|
| 39 |
+
"output_logits": false,
|
| 40 |
+
"pp_degree": 1,
|
| 41 |
+
"sequence_length": 4096,
|
| 42 |
+
"speculation_length": 0,
|
| 43 |
+
"start_rank_id": 0,
|
| 44 |
+
"target": "trn2",
|
| 45 |
+
"torch_dtype": "bfloat16",
|
| 46 |
+
"tp_degree": 4
|
| 47 |
+
},
|
| 48 |
+
"num_attention_heads": 32,
|
| 49 |
+
"num_hidden_layers": 40,
|
| 50 |
+
"num_key_value_heads": 8,
|
| 51 |
+
"residual_multiplier": 0.22,
|
| 52 |
+
"rms_norm_eps": 1e-05,
|
| 53 |
+
"rope_scaling": null,
|
| 54 |
+
"rope_theta": 5000000.0,
|
| 55 |
+
"tie_word_embeddings": true,
|
| 56 |
+
"use_cache": true,
|
| 57 |
+
"vocab_size": 49155
|
| 58 |
+
}
|
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
|
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:da2a6175e9066d90c893cb4de5c58a0cf100b67bdb2e5b4eb8d503b6c6313041
|
| 3 |
+
size 2344505
|
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:851f73f42c09d7dea928b5328d0071726bba708a96912d19d086ab353adcfae3
|
| 3 |
+
size 5315584
|
neuronxcc-2.21.33363.0+82129205/MODULE_749cb1da17f656c57ecd+ac10809c/wrapped_neff.hlo
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0f3a0298bc0f6851ee057688c3b3afff0643dacd989baec2cf75fb192f83c888
|
| 3 |
+
size 5503126
|
neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn2", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=2", "--logfile=/tmp/nxd_model/context_encoding/_tp0_bk0/log-neuron-cc.txt"]
|
neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8ae3a5d16f8966c6d9009691f650b6d64cb3576c9b88a17dc0991f2b4547b7bb
|
| 3 |
+
size 2021280
|
neuronxcc-2.21.33363.0+82129205/MODULE_7a50c1d04b1b5328b107+a32116a7/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:60d4644fc355dd5578b7be30f4698c94331f64446a290c880278eafe762ba3ca
|
| 3 |
+
size 44114944
|