#0: update HF llama-2 version

tenstorrent · Nov 20, 2023 · b2e9fd9 · b2e9fd9
1 parent ea40f8d
commit b2e9fd9
Show file tree

Hide file tree

Showing 8 changed files with 33 additions and 83 deletions.
diff --git a/models/experimental/llama_old/README.md b/models/experimental/llama_old/README.md
@@ -4,13 +4,13 @@ In order to use HuggingFace Llama model we need a development version of Transfo
 
 # Use pretrained weights
 
-The weights used in the tests are downloaded from: https://huggingface.co/decapoda-research/llama-7b-hf
+The weights used in the tests are downloaded from: https://huggingface.co/baffo32/decapoda-research-llama-7B-hf
 
 How to use the weights:
     ```
     from transformers import AutoTokenizer, AutoModelForCausalLM
-    tokenizer = AutoTokenizer.from_pretrained("decapoda-research/llama-7b-hf")
-    model = AutoModelForCausalLM.from_pretrained("decapoda-research/llama-7b-hf")
+    tokenizer = AutoTokenizer.from_pretrained(hf-internal-testing/llama-tokenizer)
+    model = AutoModelForCausalLM.from_pretrained("baffo32/decapoda-research-llama-7B-hf)
     ```
 
 An issue which probably will appear is:

diff --git a/models/experimental/llama_old/tests/test_llama_attention.py b/models/experimental/llama_old/tests/test_llama_attention.py
@@ -12,12 +12,7 @@
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 from models.experimental.llama_old.llama_utils import *
-from models.utility_functions import (
-    comp_allclose,
-    comp_pcc,
-    torch_to_tt_tensor_rm,
-    tt_to_torch_tensor
-)
+from models.utility_functions import comp_allclose, comp_pcc, torch_to_tt_tensor_rm, tt_to_torch_tensor
 
 from models.experimental.llama_old.tt.llama_attention import TtLlamaAttention
 
@@ -35,16 +30,12 @@ def forward(self, x, y):
         return result
 
 
-def run_test_LlamaAttention_inference(
-    device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-):
+def run_test_LlamaAttention_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc):
     model_name = model_version
     tokenizer_name = tokenizer_version
 
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
 
     configuration = hugging_face_reference_model.config
@@ -72,9 +63,7 @@ def run_test_LlamaAttention_inference(
     position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
 
     # PyTorch output =======================================================================
-    pytorch_LlamaAttention_model = PytorchLlamaAttentionModel(
-        hugging_face_reference_model, layer_num
-    )
+    pytorch_LlamaAttention_model = PytorchLlamaAttentionModel(hugging_face_reference_model, layer_num)
     pytorch_out = pytorch_LlamaAttention_model(x=attention_input, y=position_ids)
 
     # TT hardware execution =================================================================
@@ -114,7 +103,7 @@ def run_test_LlamaAttention_inference(
     "model_version, tokenizer_version, batch, seq_len, on_weka, pcc",
     (
         pytest.param(
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             1,
             128,
@@ -123,9 +112,5 @@ def run_test_LlamaAttention_inference(
         ),
     ),
 )
-def test_LlamaAttention_inference(
-    device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-):
-    run_test_LlamaAttention_inference(
-        device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-    )
+def test_LlamaAttention_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc):
+    run_test_LlamaAttention_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc)
diff --git a/models/experimental/llama_old/tests/test_llama_causallm.py b/models/experimental/llama_old/tests/test_llama_causallm.py
@@ -35,9 +35,7 @@ def run_test_llamaCausallm_inference(
     tokenizer_name = tokenizer_version
 
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
     configuration = hugging_face_reference_model.config
     state_dict = hugging_face_reference_model.state_dict()
@@ -91,7 +89,7 @@ def run_test_llamaCausallm_inference(
     "model_version, tokenizer_version, batch, seq_len, num_decoders, max_position_embeddings, on_weka, pcc",
     (
         (
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             4,
             128,

diff --git a/models/experimental/llama_old/tests/test_llama_decoder.py b/models/experimental/llama_old/tests/test_llama_decoder.py
@@ -9,7 +9,6 @@
 from torch import nn
 
 
-
 from typing import Tuple
 
 from transformers import AutoTokenizer, AutoModelForCausalLM
@@ -45,9 +44,7 @@ def run_test_LlamaDecoder_inference(
     tokenizer_name = tokenizer_version
 
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
     configuration = hugging_face_reference_model.config
     state_dict = hugging_face_reference_model.state_dict()
@@ -73,9 +70,7 @@ def run_test_LlamaDecoder_inference(
     position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
 
     # PyTorch output =======================================================================
-    pytorch_LlamaDecoder_model = PytorchLlamaDecoderModel(
-        hugging_face_reference_model, decoder_id
-    )
+    pytorch_LlamaDecoder_model = PytorchLlamaDecoderModel(hugging_face_reference_model, decoder_id)
     pytorch_LlamaDecoder_model.eval()
     pytorch_out = pytorch_LlamaDecoder_model(x=llama_input, y=position_ids)
 
@@ -92,9 +87,7 @@ def run_test_LlamaDecoder_inference(
         max_position_embeddings,
         configuration,
     )
-    tt_out = tt_LlamaDecoder_model(
-        hidden_states=tt_llama_input, position_ids=position_ids
-    )
+    tt_out = tt_LlamaDecoder_model(hidden_states=tt_llama_input, position_ids=position_ids)
     # transform to PyTorch tensor
     # take only hidden_states tensor if tuple is obtained
     tt_out = tt_to_torch_tensor(tt_out[0])
@@ -117,7 +110,7 @@ def run_test_LlamaDecoder_inference(
     "model_version, tokenizer_version, batch, seq_len, decoder_id, on_weka, pcc",
     (
         (
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             1,
             128,
@@ -127,9 +120,7 @@ def run_test_LlamaDecoder_inference(
         ),
     ),
 )
-def test_LlamaDecoder_inference(
-    device, model_version, tokenizer_version, batch, seq_len, decoder_id, on_weka, pcc
-):
+def test_LlamaDecoder_inference(device, model_version, tokenizer_version, batch, seq_len, decoder_id, on_weka, pcc):
     run_test_LlamaDecoder_inference(
         device,
         model_version,

diff --git a/models/experimental/llama_old/tests/test_llama_layer_norm.py b/models/experimental/llama_old/tests/test_llama_layer_norm.py
@@ -35,17 +35,13 @@ def forward(self, x):
         return result
 
 
-def run_test_LlamaLayerNorm_inference(
-    device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-):
+def run_test_LlamaLayerNorm_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc):
     model_name = model_version
     tokenizer_name = tokenizer_version
 
-    # https://huggingface.co/decapoda-research/llama-7b-hf
+    # https://huggingface.co/baffo32/decapoda-research-llama-7B-hf
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
     configuration = hugging_face_reference_model.config
     state_dict = hugging_face_reference_model.state_dict()
@@ -56,9 +52,7 @@ def run_test_LlamaLayerNorm_inference(
     layer_num = 0
 
     # PyTorch output ---------------------------------------------------------------------
-    pytorch_LlamaRMSNorm_model = PytorchLlamaRMSNormModel(
-        hugging_face_reference_model, layer_num
-    )
+    pytorch_LlamaRMSNorm_model = PytorchLlamaRMSNormModel(hugging_face_reference_model, layer_num)
     pytorch_out = pytorch_LlamaRMSNorm_model(llama_layer_norm_input)
     logger.info(f"PyTorch output shape: {pytorch_out.shape}")
 
@@ -98,7 +92,7 @@ def run_test_LlamaLayerNorm_inference(
     "model_version, tokenizer_version, batch, seq_len, on_weka, pcc",
     (
         (
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             1,
             2048,
@@ -107,9 +101,5 @@ def run_test_LlamaLayerNorm_inference(
         ),
     ),
 )
-def test_LlamaLayerNorm_inference(
-    model_version, tokenizer_version, batch, seq_len, on_weka, pcc, device
-):
-    run_test_LlamaLayerNorm_inference(
-        device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-    )
+def test_LlamaLayerNorm_inference(model_version, tokenizer_version, batch, seq_len, on_weka, pcc, device):
+    run_test_LlamaLayerNorm_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc)
diff --git a/models/experimental/llama_old/tests/test_llama_mlp.py b/models/experimental/llama_old/tests/test_llama_mlp.py
@@ -8,7 +8,6 @@
 from torch import nn
 
 
-
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
 
@@ -35,16 +34,12 @@ def forward(self, x):
         return result
 
 
-def run_test_LlamaMLP_inference(
-    device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-):
+def run_test_LlamaMLP_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc):
     model_name = model_version
     tokenizer_name = tokenizer_version
 
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
     configuration = hugging_face_reference_model.config
     state_dict = hugging_face_reference_model.state_dict()
@@ -56,9 +51,7 @@ def run_test_LlamaMLP_inference(
     base_url = "model.layers"
 
     # PyTorch output --------------------------------------------------------------------
-    pytorch_LlamaMLP_model = PytorchLlamaMLPModel(
-        hugging_face_reference_model, layer_num
-    )
+    pytorch_LlamaMLP_model = PytorchLlamaMLPModel(hugging_face_reference_model, layer_num)
     pytorch_out = pytorch_LlamaMLP_model(llama_mlp_input)  # .unsqueeze(1)
 
     # TT hardware execution -------------------------------------------------------------
@@ -94,7 +87,7 @@ def run_test_LlamaMLP_inference(
     "model_version, tokenizer_version, batch, seq_len, on_weka, pcc",
     (
         (
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             1,
             2048,
@@ -103,9 +96,5 @@ def run_test_LlamaMLP_inference(
         ),
     ),
 )
-def test_LlamaMLP_inference(
-    device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-):
-    run_test_LlamaMLP_inference(
-        device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc
-    )
+def test_LlamaMLP_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc):
+    run_test_LlamaMLP_inference(device, model_version, tokenizer_version, batch, seq_len, on_weka, pcc)
diff --git a/models/experimental/llama_old/tests/test_llama_model.py b/models/experimental/llama_old/tests/test_llama_model.py
@@ -7,7 +7,6 @@
 import pytest
 
 
-
 from models.experimental.llama.llama_utils import *
 
 from transformers import (
@@ -38,9 +37,7 @@ def run_test_Llama_inference(
     tokenizer_name = tokenizer_version
 
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(
-        model_name, torch_dtype=torch.float32
-    )
+    hugging_face_reference_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32)
     hugging_face_reference_model.eval()
     configuration = hugging_face_reference_model.config
     state_dict = hugging_face_reference_model.state_dict()
@@ -98,7 +95,7 @@ def run_test_Llama_inference(
     "model_version, tokenizer_version, batch, seq_len, num_decoders, max_position_embeddings, on_weka, pcc",
     (
         (
-            "decapoda-research/llama-7b-hf",
+            "baffo32/decapoda-research-llama-7B-hf",
             "hf-internal-testing/llama-tokenizer",
             1,
             64,

diff --git a/models/experimental/llama_old/tests/test_perf_llama.py b/models/experimental/llama_old/tests/test_perf_llama.py
@@ -25,7 +25,7 @@
 
 
 def run_perf_llama(expected_inference_time, expected_compile_time, device):
-    model_version = "decapoda-research/llama-7b-hf"
+    model_version = "baffo32/decapoda-research-llama-7B-hf"
     tokenizer_version = "hf-internal-testing/llama-tokenizer"
     batch = 1
     seq_len = 64