Synchronizing local compiler cache.
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .gitattributes +53 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/llamafactory/tiny-random-Llama-3/ae5eaaedd9d157341dd7.json +62 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/llamafactory/tiny-random-Llama-3/c8b595206cf08a8d6e71.json +62 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen2/Qwen/Qwen2.5-0.5B/b96429cfd6ed81a9a296.json +82 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen2/Qwen/Qwen2.5-0.5B/c299f278aadacee2a974.json +82 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/610feb1bc88fd055e9c7.json +87 -0
- neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/97dff437370337281482.json +87 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.neff +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/wrapped_neff.hlo +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.done +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.hlo_module.pb +3 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.neff +0 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_1210392327607194823+e30acd3a/compile_flags.json +1 -0
- neuronxcc-2.21.18209.0+043b1bf7/MODULE_1210392327607194823+e30acd3a/model.done +0 -0
.gitattributes
CHANGED
|
@@ -12928,3 +12928,56 @@ neuronxcc-2.21.18209.0+043b1bf7/MODULE_f34dea1dd0e39d38d79e+24129607/model.neff
|
|
| 12928 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_07e856db17444803b299+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12929 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_07e856db17444803b299+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12930 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b3156313afcf6f9c1c2b+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12928 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_07e856db17444803b299+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12929 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_07e856db17444803b299+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12930 |
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b3156313afcf6f9c1c2b+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12931 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12932 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12933 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_155d3059a636dbde2772+fb4cc044/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12934 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1752954cd84d208a194c+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12935 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1752954cd84d208a194c+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12936 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1df28a40cf4145bdfcae+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12937 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_254c67a8dd8bfb2e240d+ac10809c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12938 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_254c67a8dd8bfb2e240d+ac10809c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12939 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3b2b11a32ab8b63d693e+d52a1617/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12940 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3b2b11a32ab8b63d693e+d52a1617/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12941 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_3e60ddce17a65ac4ee68+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12942 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_410ef55a3aca5cd80fc8+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12943 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_54e3a5446bba5bfa79af+fb4cc044/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12944 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_57429a2762d689aab6fb+91970973/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12945 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_57429a2762d689aab6fb+91970973/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12946 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12947 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_62abb2454c6da8d57ad5+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12948 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_733744f5680ca67f7ff9+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12949 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12950 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_73587bc7122ba5a81758+80d05c3f/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12951 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_75329916c4da1120954c+6e4949b4/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12952 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_7c590ab98f45f8a8dd15+283df001/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12953 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_7c9d7dd1d913b7d0da1a+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12954 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_859cc62b15eb336cbb27+2fe092d7/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12955 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_859cc62b15eb336cbb27+2fe092d7/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12956 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_8cc67d6573976e582c95+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12957 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_913d465d9ff4ee2a67a8+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12958 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_913d465d9ff4ee2a67a8+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12959 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_944763b36d810f4af582+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12960 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_99a19dff67d566a541e8+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12961 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_99a19dff67d566a541e8+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12962 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12963 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b23271513f9095bde4db+2dde74c7/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12964 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b3b32d97fcafb37214de+09275fbf/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12965 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b3b32d97fcafb37214de+09275fbf/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12966 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b5dfe3f2e102fa0dc843+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12967 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b5dfe3f2e102fa0dc843+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12968 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_b6ed8374075f070c66d6+a32116a7/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12969 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12970 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_bdf25cde1a48d420319c+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12971 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c30a746238d3804f9626+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12972 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c4737893663f5ff66b53+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12973 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c6abc85636846983f4d9+5520a857/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12974 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_c6abc85636846983f4d9+5520a857/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12975 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12976 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_cde3417e6918fabff4a7+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12977 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d30bb77760f1638377ba+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12978 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_d30bb77760f1638377ba+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12979 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_da11c0d53c64138990e8+c43aa9be/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12980 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_da11c0d53c64138990e8+c43aa9be/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 12981 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_f52fe8e88daf6dc9619a+24129607/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12982 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 12983 |
+
neuronxcc-2.21.18209.0+043b1bf7/MODULE_fef9240b8fbfaca895ba+a02c3a36/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/llamafactory/tiny-random-Llama-3/ae5eaaedd9d157341dd7.json
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "llamafactory/tiny-random-Llama-3",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"LlamaForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "float16",
|
| 11 |
+
"head_dim": 4,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 16,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 64,
|
| 16 |
+
"max_position_embeddings": 131072,
|
| 17 |
+
"mlp_bias": false,
|
| 18 |
+
"model_type": "llama",
|
| 19 |
+
"neuron": {
|
| 20 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 21 |
+
"batch_size": 1,
|
| 22 |
+
"capacity_factor": null,
|
| 23 |
+
"checkpoint_id": "llamafactory/tiny-random-Llama-3",
|
| 24 |
+
"checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
|
| 25 |
+
"continuous_batching": false,
|
| 26 |
+
"ep_degree": 1,
|
| 27 |
+
"fused_qkv": true,
|
| 28 |
+
"glu_mlp": true,
|
| 29 |
+
"local_ranks_size": 2,
|
| 30 |
+
"max_batch_size": 1,
|
| 31 |
+
"max_context_length": 512,
|
| 32 |
+
"max_topk": 256,
|
| 33 |
+
"n_active_tokens": 512,
|
| 34 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 35 |
+
"on_device_sampling": true,
|
| 36 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 37 |
+
"output_logits": false,
|
| 38 |
+
"pp_degree": 1,
|
| 39 |
+
"sequence_length": 512,
|
| 40 |
+
"speculation_length": 0,
|
| 41 |
+
"start_rank_id": 0,
|
| 42 |
+
"target": "trn1",
|
| 43 |
+
"torch_dtype": "float16",
|
| 44 |
+
"tp_degree": 2
|
| 45 |
+
},
|
| 46 |
+
"num_attention_heads": 4,
|
| 47 |
+
"num_hidden_layers": 2,
|
| 48 |
+
"num_key_value_heads": 4,
|
| 49 |
+
"pretraining_tp": 1,
|
| 50 |
+
"rms_norm_eps": 1e-05,
|
| 51 |
+
"rope_scaling": {
|
| 52 |
+
"factor": 8.0,
|
| 53 |
+
"high_freq_factor": 4.0,
|
| 54 |
+
"low_freq_factor": 1.0,
|
| 55 |
+
"original_max_position_embeddings": 8192,
|
| 56 |
+
"rope_type": "llama3"
|
| 57 |
+
},
|
| 58 |
+
"rope_theta": 500000.0,
|
| 59 |
+
"tie_word_embeddings": false,
|
| 60 |
+
"use_cache": true,
|
| 61 |
+
"vocab_size": 128256
|
| 62 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/llama/llamafactory/tiny-random-Llama-3/c8b595206cf08a8d6e71.json
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "llamafactory/tiny-random-Llama-3",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"LlamaForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "float16",
|
| 11 |
+
"head_dim": 4,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 16,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 64,
|
| 16 |
+
"max_position_embeddings": 131072,
|
| 17 |
+
"mlp_bias": false,
|
| 18 |
+
"model_type": "llama",
|
| 19 |
+
"neuron": {
|
| 20 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 21 |
+
"batch_size": 2,
|
| 22 |
+
"capacity_factor": null,
|
| 23 |
+
"checkpoint_id": "llamafactory/tiny-random-Llama-3",
|
| 24 |
+
"checkpoint_revision": "bf2a2e3bf199ad2ee96f02a3c00246c608db22a8",
|
| 25 |
+
"continuous_batching": true,
|
| 26 |
+
"ep_degree": 1,
|
| 27 |
+
"fused_qkv": true,
|
| 28 |
+
"glu_mlp": true,
|
| 29 |
+
"local_ranks_size": 2,
|
| 30 |
+
"max_batch_size": 2,
|
| 31 |
+
"max_context_length": 128,
|
| 32 |
+
"max_topk": 256,
|
| 33 |
+
"n_active_tokens": 128,
|
| 34 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 35 |
+
"on_device_sampling": true,
|
| 36 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 37 |
+
"output_logits": false,
|
| 38 |
+
"pp_degree": 1,
|
| 39 |
+
"sequence_length": 128,
|
| 40 |
+
"speculation_length": 0,
|
| 41 |
+
"start_rank_id": 0,
|
| 42 |
+
"target": "trn2",
|
| 43 |
+
"torch_dtype": "float16",
|
| 44 |
+
"tp_degree": 2
|
| 45 |
+
},
|
| 46 |
+
"num_attention_heads": 4,
|
| 47 |
+
"num_hidden_layers": 2,
|
| 48 |
+
"num_key_value_heads": 4,
|
| 49 |
+
"pretraining_tp": 1,
|
| 50 |
+
"rms_norm_eps": 1e-05,
|
| 51 |
+
"rope_scaling": {
|
| 52 |
+
"factor": 8.0,
|
| 53 |
+
"high_freq_factor": 4.0,
|
| 54 |
+
"low_freq_factor": 1.0,
|
| 55 |
+
"original_max_position_embeddings": 8192,
|
| 56 |
+
"rope_type": "llama3"
|
| 57 |
+
},
|
| 58 |
+
"rope_theta": 500000.0,
|
| 59 |
+
"tie_word_embeddings": false,
|
| 60 |
+
"use_cache": true,
|
| 61 |
+
"vocab_size": 128256
|
| 62 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen2/Qwen/Qwen2.5-0.5B/b96429cfd6ed81a9a296.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "Qwen/Qwen2.5-0.5B",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Qwen2ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_dropout": 0.0,
|
| 9 |
+
"dtype": "bfloat16",
|
| 10 |
+
"hidden_act": "silu",
|
| 11 |
+
"hidden_size": 896,
|
| 12 |
+
"initializer_range": 0.02,
|
| 13 |
+
"intermediate_size": 4864,
|
| 14 |
+
"layer_types": [
|
| 15 |
+
"full_attention",
|
| 16 |
+
"full_attention",
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention"
|
| 39 |
+
],
|
| 40 |
+
"max_position_embeddings": 32768,
|
| 41 |
+
"max_window_layers": 24,
|
| 42 |
+
"model_type": "qwen2",
|
| 43 |
+
"neuron": {
|
| 44 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 45 |
+
"batch_size": 2,
|
| 46 |
+
"capacity_factor": null,
|
| 47 |
+
"checkpoint_id": "Qwen/Qwen2.5-0.5B",
|
| 48 |
+
"checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
|
| 49 |
+
"continuous_batching": true,
|
| 50 |
+
"ep_degree": 1,
|
| 51 |
+
"fused_qkv": false,
|
| 52 |
+
"glu_mlp": true,
|
| 53 |
+
"local_ranks_size": 2,
|
| 54 |
+
"max_batch_size": 2,
|
| 55 |
+
"max_context_length": 1024,
|
| 56 |
+
"max_topk": 256,
|
| 57 |
+
"n_active_tokens": 1024,
|
| 58 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 59 |
+
"on_device_sampling": false,
|
| 60 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 61 |
+
"output_logits": false,
|
| 62 |
+
"pp_degree": 1,
|
| 63 |
+
"sequence_length": 1024,
|
| 64 |
+
"speculation_length": 0,
|
| 65 |
+
"start_rank_id": 0,
|
| 66 |
+
"target": "trn1",
|
| 67 |
+
"torch_dtype": "bfloat16",
|
| 68 |
+
"tp_degree": 2
|
| 69 |
+
},
|
| 70 |
+
"num_attention_heads": 14,
|
| 71 |
+
"num_hidden_layers": 24,
|
| 72 |
+
"num_key_value_heads": 2,
|
| 73 |
+
"rms_norm_eps": 1e-06,
|
| 74 |
+
"rope_scaling": null,
|
| 75 |
+
"rope_theta": 1000000.0,
|
| 76 |
+
"sliding_window": null,
|
| 77 |
+
"tie_word_embeddings": true,
|
| 78 |
+
"use_cache": true,
|
| 79 |
+
"use_mrope": false,
|
| 80 |
+
"use_sliding_window": false,
|
| 81 |
+
"vocab_size": 151936
|
| 82 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen2/Qwen/Qwen2.5-0.5B/c299f278aadacee2a974.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "Qwen/Qwen2.5-0.5B",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Qwen2ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_dropout": 0.0,
|
| 9 |
+
"dtype": "bfloat16",
|
| 10 |
+
"hidden_act": "silu",
|
| 11 |
+
"hidden_size": 896,
|
| 12 |
+
"initializer_range": 0.02,
|
| 13 |
+
"intermediate_size": 4864,
|
| 14 |
+
"layer_types": [
|
| 15 |
+
"full_attention",
|
| 16 |
+
"full_attention",
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention"
|
| 39 |
+
],
|
| 40 |
+
"max_position_embeddings": 32768,
|
| 41 |
+
"max_window_layers": 24,
|
| 42 |
+
"model_type": "qwen2",
|
| 43 |
+
"neuron": {
|
| 44 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 45 |
+
"batch_size": 1,
|
| 46 |
+
"capacity_factor": null,
|
| 47 |
+
"checkpoint_id": "Qwen/Qwen2.5-0.5B",
|
| 48 |
+
"checkpoint_revision": "060db6499f32faf8b98477b0a26969ef7d8b9987",
|
| 49 |
+
"continuous_batching": false,
|
| 50 |
+
"ep_degree": 1,
|
| 51 |
+
"fused_qkv": false,
|
| 52 |
+
"glu_mlp": true,
|
| 53 |
+
"local_ranks_size": 2,
|
| 54 |
+
"max_batch_size": 1,
|
| 55 |
+
"max_context_length": 4096,
|
| 56 |
+
"max_topk": 256,
|
| 57 |
+
"n_active_tokens": 4096,
|
| 58 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 59 |
+
"on_device_sampling": true,
|
| 60 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 61 |
+
"output_logits": false,
|
| 62 |
+
"pp_degree": 1,
|
| 63 |
+
"sequence_length": 4096,
|
| 64 |
+
"speculation_length": 0,
|
| 65 |
+
"start_rank_id": 0,
|
| 66 |
+
"target": "trn1",
|
| 67 |
+
"torch_dtype": "bfloat16",
|
| 68 |
+
"tp_degree": 2
|
| 69 |
+
},
|
| 70 |
+
"num_attention_heads": 14,
|
| 71 |
+
"num_hidden_layers": 24,
|
| 72 |
+
"num_key_value_heads": 2,
|
| 73 |
+
"rms_norm_eps": 1e-06,
|
| 74 |
+
"rope_scaling": null,
|
| 75 |
+
"rope_theta": 1000000.0,
|
| 76 |
+
"sliding_window": null,
|
| 77 |
+
"tie_word_embeddings": true,
|
| 78 |
+
"use_cache": true,
|
| 79 |
+
"use_mrope": false,
|
| 80 |
+
"use_sliding_window": false,
|
| 81 |
+
"vocab_size": 151936
|
| 82 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/610feb1bc88fd055e9c7.json
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "Qwen/Qwen3-Embedding-0.6B",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Qwen3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"head_dim": 128,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 1024,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 3072,
|
| 16 |
+
"layer_types": [
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention"
|
| 45 |
+
],
|
| 46 |
+
"max_position_embeddings": 32768,
|
| 47 |
+
"max_window_layers": 28,
|
| 48 |
+
"model_type": "qwen3",
|
| 49 |
+
"neuron": {
|
| 50 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 51 |
+
"batch_size": 6,
|
| 52 |
+
"capacity_factor": null,
|
| 53 |
+
"checkpoint_id": "Qwen/Qwen3-Embedding-0.6B",
|
| 54 |
+
"checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418",
|
| 55 |
+
"continuous_batching": false,
|
| 56 |
+
"ep_degree": 1,
|
| 57 |
+
"fused_qkv": true,
|
| 58 |
+
"glu_mlp": true,
|
| 59 |
+
"local_ranks_size": 2,
|
| 60 |
+
"max_batch_size": 6,
|
| 61 |
+
"max_context_length": 1024,
|
| 62 |
+
"max_topk": 256,
|
| 63 |
+
"n_active_tokens": 1024,
|
| 64 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 65 |
+
"on_device_sampling": false,
|
| 66 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 67 |
+
"output_logits": false,
|
| 68 |
+
"pp_degree": 1,
|
| 69 |
+
"sequence_length": 1024,
|
| 70 |
+
"speculation_length": 0,
|
| 71 |
+
"start_rank_id": 0,
|
| 72 |
+
"target": "trn1",
|
| 73 |
+
"torch_dtype": "bfloat16",
|
| 74 |
+
"tp_degree": 2
|
| 75 |
+
},
|
| 76 |
+
"num_attention_heads": 16,
|
| 77 |
+
"num_hidden_layers": 28,
|
| 78 |
+
"num_key_value_heads": 8,
|
| 79 |
+
"rms_norm_eps": 1e-06,
|
| 80 |
+
"rope_scaling": null,
|
| 81 |
+
"rope_theta": 1000000,
|
| 82 |
+
"sliding_window": null,
|
| 83 |
+
"tie_word_embeddings": true,
|
| 84 |
+
"use_cache": true,
|
| 85 |
+
"use_sliding_window": false,
|
| 86 |
+
"vocab_size": 151669
|
| 87 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/0_REGISTRY/0.4.2.dev2/qwen3/Qwen/Qwen3-Embedding-0.6B/97dff437370337281482.json
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "Qwen/Qwen3-Embedding-0.6B",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"Qwen3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": false,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"dtype": "bfloat16",
|
| 11 |
+
"head_dim": 128,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 1024,
|
| 14 |
+
"initializer_range": 0.02,
|
| 15 |
+
"intermediate_size": 3072,
|
| 16 |
+
"layer_types": [
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention"
|
| 45 |
+
],
|
| 46 |
+
"max_position_embeddings": 32768,
|
| 47 |
+
"max_window_layers": 28,
|
| 48 |
+
"model_type": "qwen3",
|
| 49 |
+
"neuron": {
|
| 50 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 51 |
+
"batch_size": 4,
|
| 52 |
+
"capacity_factor": null,
|
| 53 |
+
"checkpoint_id": "Qwen/Qwen3-Embedding-0.6B",
|
| 54 |
+
"checkpoint_revision": "c54f2e6e80b2d7b7de06f51cec4959f6b3e03418",
|
| 55 |
+
"continuous_batching": false,
|
| 56 |
+
"ep_degree": 1,
|
| 57 |
+
"fused_qkv": true,
|
| 58 |
+
"glu_mlp": true,
|
| 59 |
+
"local_ranks_size": 2,
|
| 60 |
+
"max_batch_size": 4,
|
| 61 |
+
"max_context_length": 1024,
|
| 62 |
+
"max_topk": 256,
|
| 63 |
+
"n_active_tokens": 1024,
|
| 64 |
+
"neuronxcc_version": "2.21.18209.0+043b1bf7",
|
| 65 |
+
"on_device_sampling": false,
|
| 66 |
+
"optimum_neuron_version": "0.4.2.dev2",
|
| 67 |
+
"output_logits": false,
|
| 68 |
+
"pp_degree": 1,
|
| 69 |
+
"sequence_length": 1024,
|
| 70 |
+
"speculation_length": 0,
|
| 71 |
+
"start_rank_id": 0,
|
| 72 |
+
"target": "trn1",
|
| 73 |
+
"torch_dtype": "bfloat16",
|
| 74 |
+
"tp_degree": 2
|
| 75 |
+
},
|
| 76 |
+
"num_attention_heads": 16,
|
| 77 |
+
"num_hidden_layers": 28,
|
| 78 |
+
"num_key_value_heads": 8,
|
| 79 |
+
"rms_norm_eps": 1e-06,
|
| 80 |
+
"rope_scaling": null,
|
| 81 |
+
"rope_theta": 1000000,
|
| 82 |
+
"sliding_window": null,
|
| 83 |
+
"tie_word_embeddings": true,
|
| 84 |
+
"use_cache": true,
|
| 85 |
+
"use_sliding_window": false,
|
| 86 |
+
"vocab_size": 151669
|
| 87 |
+
}
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1", "--auto-cast=none", "--model-type=transformer", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2 --vectorize-strided-dma ", "-O2", "--lnc=1", "--logfile=/tmp/nxd_model/token_generation/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f0c908a5ddbf03970315b73168cfd6d3b14f39328f2b700c5d1e9c4144323034
|
| 3 |
+
size 91147
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7cefc6bcc07baf397dce345288c5b77016de342527ed081816143e4ce55a8892
|
| 3 |
+
size 277504
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_0cdc152aff6a7cbf36f6+a02c3a36/wrapped_neff.hlo
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f6a16f4ffabef1dd579e775328d0503dd78c2fd022be2edbc934fbd4dbc57be0
|
| 3 |
+
size 289031
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:68d800848ea22ac4d6f5cf0e1e6088da78460e16b5340dd0f956b5b5bf0c1834
|
| 3 |
+
size 1122
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10217061096959125489+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0113611968c8ed0cbeaea5ff3bba2ce4eecd5fc989e2592c6b8ac5fdf09c91c2
|
| 3 |
+
size 1562
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10244305442015770634+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a35335d3e96ee86956cdf93862bf42ae6f83017e9b331b77531ad6942a62366d
|
| 3 |
+
size 1269
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10606948783918825529+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:153fdddade7e1b284f8d78df84523bd332eccc7ddfc7a0698ad8653d560c55ca
|
| 3 |
+
size 1124
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10645643398657092095+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7434cf1ede5dc6759a8ad5cdc472fb8e1668b1a8bd8dd7e7e08b5c09401b8065
|
| 3 |
+
size 1562
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10661660426924300837+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e7000c9195e6c3c9abfcbac09928468ef4c204488094a046ec7abfd1b7ea8657
|
| 3 |
+
size 1563
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_10746122569655005679+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2fcfd4f1816e23bee263009b76a032d3de2454e4937f73423fade31c7a0f0278
|
| 3 |
+
size 1562
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11086318750207148626+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:701e22032f5900dff59bf252dc2804e14345ee4b52f9587d7305edd2a497770f
|
| 3 |
+
size 1269
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11909310600244571805+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.done
ADDED
|
File without changes
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a3ba973ff38c3c4be9f7ceda4e02dd7d04d9799bf7507445264e5a513b8bcc4e
|
| 3 |
+
size 1124
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_11933818254123612383+e30acd3a/model.neff
ADDED
|
Binary file (31.7 kB). View file
|
|
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1210392327607194823+e30acd3a/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1"]
|
neuronxcc-2.21.18209.0+043b1bf7/MODULE_1210392327607194823+e30acd3a/model.done
ADDED
|
File without changes
|