[BOUNTY] Add Phi-2 (#117)

* Qwen1.5 0.5B pybuda implementation

* remove unneeded requirement

* rename "acceleration" to "accelerate"

* Update env vars and compiler configs

* remove undefined device_map

* Remove misleading and unnecessary environment variables

* remove qwen from phi branch

* Add Phi 2

* Update requirements.txt

* Standardize Phi2 demo and added tests

* Remove old phi2 demo

* fix missing quote in pyproject.toml

* fix

* Fix test saying qwen1_5 instead of phi2

model_demos/nlp_demos/phi2/pytorch_phi2_text_generation.py

@@ -0,0 +1,68 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent AI ULC
+# SPDX-License-Identifier: Apache-2.0
+
+# Phi2 Demo - Text Generation
+
+import os
+import pybuda
+
+from transformers import PhiForCausalLM, AutoTokenizer, PhiConfig
+from pybuda.transformers.pipeline import pipeline as pybuda_pipeline
+
+def run_phi2_causal_lm(batch_size=1):
+    os.environ["TT_BACKEND_TIMEOUT"] = '0'
+
+    # Set PyBuda configurations
+    compiler_cfg = pybuda.config._get_global_compiler_config()
+    compiler_cfg.default_df_override = pybuda.DataFormat.Float16_b
+    compiler_cfg.enable_auto_fusing = True
+    compiler_cfg.balancer_policy = "Ribbon"
+
+    # Setup model configuration
+    config = PhiConfig.from_pretrained("microsoft/phi-2")
+    config.use_cache = False
+    config.return_dict = False
+
+    # Load model and tokenizer with config
+    model = PhiForCausalLM.from_pretrained("microsoft/phi-2", config=config)
+    tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")
+    tokenizer.pad_token, tokenizer.pad_token_id = (tokenizer.eos_token, tokenizer.eos_token_id)
+
+    # Disable DynamicCache
+    # See: https://github.com/tenstorrent/tt-buda/issues/42
+    model._supports_cache_class = False
+
+    # Example usage
+    prompt = ["My name is Jim Keller and"] * batch_size
+
+    # Initialize pipeline
+    text_generator = pybuda_pipeline(
+        "text-generation",
+        model=model,
+        tokenizer=tokenizer
+    )
+
+    # Inference on TT device
+    response = text_generator(
+        prompt,
+        temperature=0.7,
+        top_k=50,
+        top_p=0.9,
+        max_new_tokens=512,
+        num_beams=1,
+        do_sample=True,
+        no_repeat_ngram_size=5,
+        pad_token_id=tokenizer.pad_token_id,
+        eos_token_id=tokenizer.eos_token_id,
+        early_stopping=True
+    )
+
+    # Display Responses
+    for batch_id in range(batch_size):
+        print(f"Batch: {batch_id}")
+        print(f"Response: {response[batch_id][0]['generated_text']}")
+        print()
+
+
+if __name__ == "__main__":
+    run_phi2_causal_lm()

model_demos/pyproject.toml

@@ -87,5 +87,6 @@ markers = [
     "yolov6: tests that involve yolov6",
     "segformer: tests that involve SegFormer",
     "monodle: tests that involve Monodle",
+    "phi2: tests that involve Phi2",
     "yolox: tests that involve YOLOX",
 ]

model_demos/tests/test_pytorch_phi2.py

@@ -0,0 +1,10 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent AI ULC
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+
+from nlp_demos.phi2.pytorch_phi2_text_generation import run_phi2_causal_lm
+
+@pytest.mark.phi2
+def test_phi2_causal_lm_pytorch(clear_pybuda, test_device, batch_size):
+    run_phi2_causal_lm(batch_size=batch_size)