Add initial code for LLM to output text

Just the scaffolding for now. Showing system and user prompt
on the Streamlit app, and then a random sample button and
the LLM output response.

.gitignore

@@ -0,0 +1 @@
+__pycache__

LICENSE

@@ -0,0 +1,8 @@
+Copyright 2025-Present Justin Silver
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+

README.md

@@ -0,0 +1,21 @@
+# AutoHDL
+
+AI agent which generates Verilog code
+
+This project is a work-in-progress!
+
+## Introduction
+
+This is an attempt to build an AI agent which tries to generate syntactically correct
+Verilog code from input text specifications.
+
+When I'm done, the LLM will be able to call Verilog linter and simulation tools.
+It will then use the tools output to self-correct the Verilog code it suggests.
+
+For testing and illustration purposes, I'm using the MG-Verilog dataset from Georgia Tech.
+
+## Tools
+
+- `transformers` for LLM calls
+- `outlines` for structured LLM output
+- `streamlit` for pipeline visualization

autohdl/data.py

@@ -0,0 +1,49 @@
+import re
+from datasets import load_dataset
+
+def extract_description(prompt):
+    """Use regex to extract description from prompt
+    
+    The MG-verilog dataset comes with the special tokens such as 
+    [INST], <<SYS>>, etc. This removes those and only extracts
+    the description and verilog module header. We do this 
+    so that we can use models with different chat templates.
+    """
+    
+    sysend_re = re.compile(r"<<\/SYS>>", re.MULTILINE)
+    instend_re = re.compile(r"\[\/INST\]", re.MULTILINE)
+
+    try: 
+        sysend_pos = re.search(sysend_re, prompt).end()
+        instend_pos = re.search(instend_re, prompt).start()
+    except:
+        raise Exception("Prompt is not in expected format when extracting description")
+    
+    return prompt[sysend_pos:instend_pos].strip()
+
+def replace_template(batch):
+    """Remove system prompt and add raw description text to all summaries"""
+    
+    for batch_idx in range(len(batch['description'])):
+        descriptions = batch['description'][batch_idx]
+        
+        for summary_type in descriptions:
+            descriptions[summary_type] = extract_description(descriptions[summary_type])
+    return batch
+    
+
+def data(name="GaTech-EIC/MG-Verilog", batch_size=4, small_dataset=True):
+    """Load MG-Verilog dataset from GAtech paper
+    
+    https://arxiv.org/pdf/2407.01910
+    """
+    
+    ds = load_dataset(name, split=f"train[:{10 if small_dataset else ''}]")
+    ds = ds.map(replace_template, batched=True, batch_size=batch_size)
+    
+    return ds
+    
+
+if __name__ == "__main__":
+    data()
+    

autohdl/llm.py

@@ -0,0 +1,43 @@
+import outlines
+from outlines.inputs import Chat
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from .data import data
+
+system_prompt = ("You only complete chats with syntax correct Verilog code. "
+                 "End the Verilog module code completion with 'endmodule'. "
+                 "Do not include module, input and output definitions.")
+
+class LLM:
+    def __init__(self):
+        self.system_prompt = system_prompt
+        
+    def load_model(self, use_cpu=True):
+        if use_cpu:
+            model_name = "HuggingFaceTB/SmolLM2-360M-Instruct"
+            self.device = "cpu"
+        else:
+            model_name = "codellama/CodeLlama-7b-Instruct-hf"
+            self.device = "cuda"
+
+        self.hf_tokenizer = AutoTokenizer.from_pretrained(model_name)
+        self.hf_model = AutoModelForCausalLM.from_pretrained(model_name)
+        
+        return self.hf_tokenizer, self.hf_model
+
+    def __call__(self, description_prompt):
+
+        messages = [
+            {"role": "system", "content": self.system_prompt},
+            {"role": "user", "content": description_prompt}
+        ]
+
+        input_text=self.hf_tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        # print(input_text)
+        
+        inputs = self.hf_tokenizer.encode(input_text, return_tensors="pt").to(self.device)
+        
+        outputs = self.hf_model.generate(inputs)
+        
+        # print(self.hf_tokenizer.decode(outputs[0]))
+        return self.hf_tokenizer.decode(outputs[0])
+

requirements.txt

@@ -0,0 +1,4 @@
+transformers
+datasets
+outlines[transformers]
+streamlit

server.py

@@ -0,0 +1,70 @@
+import streamlit as st
+from autohdl.data import data
+from autohdl.llm import system_prompt, LLM
+import random
+
+"""
+# AutoHDL
+### AI agent which generates Verilog code
+"""
+
+def random_sample_btn(stop):
+    """Generate a new sample"""
+    
+    idx = random.randrange(stop)
+    st.session_state['idx'] = idx
+    print(idx)
+    return idx
+
+def generate_btn(model, text):
+    st.session_state['response'] = model(text)
+    return st.session_state['response']
+
+@st.cache_resource(show_spinner="Loading LLM...")
+def load_model():
+    model = LLM()
+    model.load_model()
+    return model 
+
+
+def server():
+    ds = data(small_dataset=True)
+    
+    model = load_model()
+    
+    if 'idx' not in st.session_state:
+        st.session_state['idx'] = 0
+    
+    if 'response' not in st.session_state:
+        st.session_state['response'] = "Click generate for LLM to respond"
+        
+    idx = st.session_state['idx']
+        
+    summary = "high_level_global_summary"
+    
+    st.text_area("System prompt",
+                 system_prompt,
+                 height="content")
+    
+    description_prompt = ds['description'][idx][summary]
+    
+    st.text_area("User prompt",
+                 description_prompt,
+                 height=200)
+    
+    st.text_area("Expected response",
+                 ds['code'][idx],
+                 height=200)
+    
+    st.button("Random sample", on_click=random_sample_btn, args=[ds.num_rows])
+    
+    st.text_area("LLM response",
+                 st.session_state['response'],
+                 disabled=True,
+                 height=200)
+    
+    st.button("Generate", on_click=generate_btn, args=[model, ds['description'][idx][summary]])
+    
+
+if __name__ == "__main__":
+    server()