Synchronizing local compiler cache.
Browse files- .gitattributes +4 -0
- neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev0/gpt_oss/plop-internal/gpt-oss-20b-trfs-0804/5663ce3c2c64f64092cc.json +120 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/compile_flags.json +1 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.done +0 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.hlo_module.pb +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.neff +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/wrapped_neff.hlo +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/compile_flags.json +1 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.done +0 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.hlo_module.pb +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.neff +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/compile_flags.json +1 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.done +0 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.hlo_module.pb +3 -0
- neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.neff +3 -0
.gitattributes
CHANGED
|
@@ -3920,3 +3920,7 @@ neuronxcc-2.19.8089.0+8ab9f450/MODULE_9ddafe9c53dfe7a4a0ca+ca355898/wrapped_neff
|
|
| 3920 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_7aa6e291958ce4b691a4+d859950c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 3921 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_7aa6e291958ce4b691a4+d859950c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 3922 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_b7b5848e8b5f86e50e8c+503f0791/model.neff filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3920 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_7aa6e291958ce4b691a4+d859950c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 3921 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_7aa6e291958ce4b691a4+d859950c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 3922 |
neuronxcc-2.19.8089.0+8ab9f450/MODULE_b7b5848e8b5f86e50e8c+503f0791/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 3923 |
+
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 3924 |
+
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/wrapped_neff.hlo filter=lfs diff=lfs merge=lfs -text
|
| 3925 |
+
neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.neff filter=lfs diff=lfs merge=lfs -text
|
| 3926 |
+
neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.neff filter=lfs diff=lfs merge=lfs -text
|
neuronxcc-2.19.8089.0+8ab9f450/0_REGISTRY/0.3.1.dev0/gpt_oss/plop-internal/gpt-oss-20b-trfs-0804/5663ce3c2c64f64092cc.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_entry_class": "SingleModelCacheEntry",
|
| 3 |
+
"_model_id": "plop-internal/gpt-oss-20b-trfs-0804",
|
| 4 |
+
"_task": "text-generation",
|
| 5 |
+
"architectures": [
|
| 6 |
+
"GptOssForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_bias": true,
|
| 9 |
+
"attention_dropout": 0.0,
|
| 10 |
+
"experts_per_token": 4,
|
| 11 |
+
"head_dim": 64,
|
| 12 |
+
"hidden_act": "silu",
|
| 13 |
+
"hidden_size": 2880,
|
| 14 |
+
"initial_context_length": 4096,
|
| 15 |
+
"initializer_range": 0.02,
|
| 16 |
+
"intermediate_size": 2880,
|
| 17 |
+
"layer_types": [
|
| 18 |
+
"sliding_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"sliding_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"sliding_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"sliding_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"sliding_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"sliding_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"sliding_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"sliding_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"sliding_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention"
|
| 42 |
+
],
|
| 43 |
+
"max_position_embeddings": 131072,
|
| 44 |
+
"mlp_bias": false,
|
| 45 |
+
"model_type": "gpt_oss",
|
| 46 |
+
"neuron": {
|
| 47 |
+
"_serialized_key": "NxDNeuronConfig",
|
| 48 |
+
"async_mode": false,
|
| 49 |
+
"attn_kernel_enabled": false,
|
| 50 |
+
"batch_size": 1,
|
| 51 |
+
"capacity_factor": null,
|
| 52 |
+
"cc_pipeline_tiling_factor": 2,
|
| 53 |
+
"checkpoint_id": "plop-internal/gpt-oss-20b-trfs-0804",
|
| 54 |
+
"checkpoint_revision": "8964acebadbe440e1f70a40383a95d7f9eec20bc",
|
| 55 |
+
"continuous_batching": false,
|
| 56 |
+
"enable_bucketing": false,
|
| 57 |
+
"ep_degree": 1,
|
| 58 |
+
"flash_decoding_enabled": false,
|
| 59 |
+
"fused_qkv": false,
|
| 60 |
+
"glu_mlp": true,
|
| 61 |
+
"is_chunked_prefill": false,
|
| 62 |
+
"local_ranks_size": 8,
|
| 63 |
+
"logical_nc_config": 1,
|
| 64 |
+
"max_batch_size": 1,
|
| 65 |
+
"max_context_length": 4096,
|
| 66 |
+
"max_topk": 256,
|
| 67 |
+
"mlp_kernel_enabled": false,
|
| 68 |
+
"mlp_kernel_fuse_residual_add": false,
|
| 69 |
+
"n_active_tokens": 4096,
|
| 70 |
+
"neuronxcc_version": "2.19.8089.0+8ab9f450",
|
| 71 |
+
"num_cores_per_group": 1,
|
| 72 |
+
"on_device_sampling": false,
|
| 73 |
+
"optimum_neuron_version": "0.3.1.dev0",
|
| 74 |
+
"output_logits": false,
|
| 75 |
+
"padding_side": "right",
|
| 76 |
+
"pp_degree": 1,
|
| 77 |
+
"qk_layernorm": false,
|
| 78 |
+
"qkv_kernel_enabled": false,
|
| 79 |
+
"rpl_reduce_dtype": "bfloat16",
|
| 80 |
+
"sequence_length": 4096,
|
| 81 |
+
"sequence_parallel_enabled": false,
|
| 82 |
+
"speculation_length": 0,
|
| 83 |
+
"start_rank_id": 0,
|
| 84 |
+
"target": null,
|
| 85 |
+
"torch_dtype": "bfloat16",
|
| 86 |
+
"tp_degree": 8,
|
| 87 |
+
"vocab_parallel": false
|
| 88 |
+
},
|
| 89 |
+
"num_attention_heads": 64,
|
| 90 |
+
"num_experts_per_tok": 4,
|
| 91 |
+
"num_hidden_layers": 1,
|
| 92 |
+
"num_key_value_heads": 8,
|
| 93 |
+
"num_local_experts": 32,
|
| 94 |
+
"output_router_logits": false,
|
| 95 |
+
"quantization_config": {
|
| 96 |
+
"modules_to_not_convert": [
|
| 97 |
+
"model.layers.*.self_attn",
|
| 98 |
+
"model.layers.*.mlp.router",
|
| 99 |
+
"model.embed_tokens",
|
| 100 |
+
"lm_head"
|
| 101 |
+
],
|
| 102 |
+
"quant_method": "mxfp4"
|
| 103 |
+
},
|
| 104 |
+
"rms_norm_eps": 1e-05,
|
| 105 |
+
"rope_scaling": {
|
| 106 |
+
"beta_fast": 32.0,
|
| 107 |
+
"beta_slow": 1.0,
|
| 108 |
+
"factor": 32.0,
|
| 109 |
+
"original_max_position_embeddings": 4096,
|
| 110 |
+
"rope_type": "yarn",
|
| 111 |
+
"truncate": false
|
| 112 |
+
},
|
| 113 |
+
"rope_theta": 150000,
|
| 114 |
+
"router_aux_loss_coef": 0.9,
|
| 115 |
+
"sliding_window": 128,
|
| 116 |
+
"swiglu_limit": 7.0,
|
| 117 |
+
"tie_word_embeddings": false,
|
| 118 |
+
"use_cache": true,
|
| 119 |
+
"vocab_size": 201088
|
| 120 |
+
}
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O0", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--logfile=/tmp/nxd_model/token_generation_model/_tp0_bk0/log-neuron-cc.txt", "--enable-internal-neff-wrapper"]
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.done
ADDED
|
File without changes
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:993d608905679fd8cc0584ab0909aa296a0d59255cfd448cabdefe38746eba4a
|
| 3 |
+
size 49245
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:15847193b5b8cb700d88f4412361358c3af736ad27b525e3f03d062b5d49ad3f
|
| 3 |
+
size 646144
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_734d88d62b55d9271aad+d859950c/wrapped_neff.hlo
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9bf9b5b6e004a0bc52c78cc0e2f8e920e2ab57924a2e8c1124ca65220a2b39ee
|
| 3 |
+
size 653119
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
"--model-type=transformer -O1 --lnc=1 --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --logfile=/tmp/nxd_model/layout_opt/log-neuron-cc.txt"
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.done
ADDED
|
File without changes
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bcdb782afaef94c23bef93d96917de7e7714161d4694f7da03d5322a8f8444c0
|
| 3 |
+
size 7011
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_aff1b3d88ed34435f1d7+431f5505/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:99c45fa0baafac8b4b6dc172c5ea1bb7274e8a95317378a1152463c6e094c1a4
|
| 3 |
+
size 144384
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/compile_flags.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["--target=trn1", "--enable-saturate-infinity", "--enable-mixed-precision-accumulation", "--model-type", "transformer", "-O0", "--tensorizer-options=--enable-ccop-compute-overlap --cc-pipeline-tiling-factor=2", "--auto-cast=none", "--internal-enable-dge-levels", "vector_dynamic_offsets", "--logfile=/tmp/nxd_model/context_encoding_model/_tp0_bk0/log-neuron-cc.txt"]
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.done
ADDED
|
File without changes
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.hlo_module.pb
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b3d3c536e9816f5383e6376a7e5cbc50acc1acb57e407493b79a471d511a5bdd
|
| 3 |
+
size 17667741
|
neuronxcc-2.19.8089.0+8ab9f450/MODULE_de16071d3a627c984e64+503f0791/model.neff
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:00a36b0465678a5d8e73f549072c86f8877b9bfedbd42e7577e5802f0adcc9d1
|
| 3 |
+
size 8510464
|