Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- README.md +30 -30
- inference.py +5 -5
- server/compiler_env_environment.py +39 -15
README.md
CHANGED
|
@@ -13,7 +13,7 @@ tags:
|
|
| 13 |
|
| 14 |
# IRis Compiler Optimization Environment
|
| 15 |
|
| 16 |
-
An OpenEnv environment where an **LLM agent selects LLVM optimization passes** to minimize execution time of C programs compiled for **RISC-V architecture**. The agent builds a sequence of compiler passes, then compiles and measures performance against standard optimization level baselines (O0
|
| 17 |
|
| 18 |
## Quick Start
|
| 19 |
|
|
@@ -21,7 +21,7 @@ An OpenEnv environment where an **LLM agent selects LLVM optimization passes** t
|
|
| 21 |
from compiler_env import CompilerAction, CompilerEnv
|
| 22 |
|
| 23 |
with CompilerEnv.from_docker_image("compiler_env-env:latest") as env:
|
| 24 |
-
# Start episode
|
| 25 |
result = env.reset()
|
| 26 |
print(f"Program: {result.observation.data['program']}")
|
| 27 |
print(f"O3 baseline: {result.observation.data['baselines']['O3']}s")
|
|
@@ -40,9 +40,9 @@ with CompilerEnv.from_docker_image("compiler_env-env:latest") as env:
|
|
| 40 |
## How It Works
|
| 41 |
|
| 42 |
1. **reset()** picks a C program (random from 200 training programs, or a fixed eval task)
|
| 43 |
-
2. The environment compiles it with `-O0`, `-O1`, `-O2`, `-O3` and measures each on QEMU
|
| 44 |
3. The agent calls tools to build an optimization pass sequence
|
| 45 |
-
4. **compile_and_measure** runs the full pipeline: `clang
|
| 46 |
5. Reward is based on how the agent's execution time compares to baselines
|
| 47 |
|
| 48 |
## Action Space
|
|
@@ -58,11 +58,11 @@ with CompilerEnv.from_docker_image("compiler_env-env:latest") as env:
|
|
| 58 |
## Observation
|
| 59 |
|
| 60 |
**CompilerObservation** fields:
|
| 61 |
-
- **data** (dict)
|
| 62 |
-
- **status** (str)
|
| 63 |
-
- **message** (str)
|
| 64 |
-
- **reward** (float)
|
| 65 |
-
- **done** (bool)
|
| 66 |
|
| 67 |
## Reward Structure (Partial Progress)
|
| 68 |
|
|
@@ -88,9 +88,9 @@ with CompilerEnv.from_docker_image("compiler_env-env:latest") as env:
|
|
| 88 |
## System Requirements
|
| 89 |
|
| 90 |
The Docker image includes all needed tools:
|
| 91 |
-
- **clang** + **llvm** (opt, llc)
|
| 92 |
-
- **gcc-riscv64-linux-gnu**
|
| 93 |
-
- **qemu-user** (qemu-riscv64)
|
| 94 |
|
| 95 |
## Building & Deploying
|
| 96 |
|
|
@@ -113,22 +113,22 @@ uvicorn server.app:app --reload --host 0.0.0.0 --port 8000
|
|
| 113 |
|
| 114 |
```
|
| 115 |
compiler_env/
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
```
|
|
|
|
| 13 |
|
| 14 |
# IRis Compiler Optimization Environment
|
| 15 |
|
| 16 |
+
An OpenEnv environment where an **LLM agent selects LLVM optimization passes** to minimize execution time of C programs compiled for **RISC-V architecture**. The agent builds a sequence of compiler passes, then compiles and measures performance against standard optimization level baselines (O0-O3).
|
| 17 |
|
| 18 |
## Quick Start
|
| 19 |
|
|
|
|
| 21 |
from compiler_env import CompilerAction, CompilerEnv
|
| 22 |
|
| 23 |
with CompilerEnv.from_docker_image("compiler_env-env:latest") as env:
|
| 24 |
+
# Start episode - picks a C program and computes baselines
|
| 25 |
result = env.reset()
|
| 26 |
print(f"Program: {result.observation.data['program']}")
|
| 27 |
print(f"O3 baseline: {result.observation.data['baselines']['O3']}s")
|
|
|
|
| 40 |
## How It Works
|
| 41 |
|
| 42 |
1. **reset()** picks a C program (random from 200 training programs, or a fixed eval task)
|
| 43 |
+
2. The environment compiles it with `-O0`, `-O1`, `-O2`, `-O3` and measures each on QEMU -> **baselines**
|
| 44 |
3. The agent calls tools to build an optimization pass sequence
|
| 45 |
+
4. **compile_and_measure** runs the full pipeline: `clang -> opt (agent's passes) -> llc -> gcc -> qemu`
|
| 46 |
5. Reward is based on how the agent's execution time compares to baselines
|
| 47 |
|
| 48 |
## Action Space
|
|
|
|
| 58 |
## Observation
|
| 59 |
|
| 60 |
**CompilerObservation** fields:
|
| 61 |
+
- **data** (dict) - Action-specific data (baselines, sequences, execution times, etc.)
|
| 62 |
+
- **status** (str) - `"success"`, `"error"`, `"invalid_pass"`, `"failed"`
|
| 63 |
+
- **message** (str) - Human-readable description
|
| 64 |
+
- **reward** (float) - Cumulative episode reward
|
| 65 |
+
- **done** (bool) - Whether the episode has ended
|
| 66 |
|
| 67 |
## Reward Structure (Partial Progress)
|
| 68 |
|
|
|
|
| 88 |
## System Requirements
|
| 89 |
|
| 90 |
The Docker image includes all needed tools:
|
| 91 |
+
- **clang** + **llvm** (opt, llc) - LLVM compilation toolchain
|
| 92 |
+
- **gcc-riscv64-linux-gnu** - RISC-V cross-compiler
|
| 93 |
+
- **qemu-user** (qemu-riscv64) - RISC-V binary emulation
|
| 94 |
|
| 95 |
## Building & Deploying
|
| 96 |
|
|
|
|
| 113 |
|
| 114 |
```
|
| 115 |
compiler_env/
|
| 116 |
+
|-- __init__.py # Module exports
|
| 117 |
+
|-- README.md # This file
|
| 118 |
+
|-- openenv.yaml # OpenEnv manifest
|
| 119 |
+
|-- pyproject.toml # Dependencies
|
| 120 |
+
|-- client.py # CompilerEnv WebSocket client
|
| 121 |
+
|-- models.py # CompilerAction & CompilerObservation
|
| 122 |
+
|-- inference.py # Baseline script (OpenAI API)
|
| 123 |
+
|-- training_programs/ # 200 C programs
|
| 124 |
+
| |-- 01_insertion_sort.c
|
| 125 |
+
| |-- 08_strassen_matrix.c
|
| 126 |
+
| |-- 114_polynomial_multiply_fft.c
|
| 127 |
+
| +-- ... (197 more)
|
| 128 |
+
+-- server/
|
| 129 |
+
|-- __init__.py
|
| 130 |
+
|-- compiler_env_environment.py # Core environment logic
|
| 131 |
+
|-- app.py # FastAPI app (HTTP + WebSocket)
|
| 132 |
+
|-- requirements.txt
|
| 133 |
+
+-- Dockerfile
|
| 134 |
```
|
inference.py
CHANGED
|
@@ -45,8 +45,8 @@ from openai import OpenAI
|
|
| 45 |
|
| 46 |
from compiler_env import CompilerAction, CompilerEnv
|
| 47 |
|
| 48 |
-
|
| 49 |
-
|
| 50 |
|
| 51 |
API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
|
| 52 |
MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"
|
|
@@ -189,9 +189,9 @@ def get_model_tool_call(client: OpenAI, messages: list) -> tuple:
|
|
| 189 |
|
| 190 |
|
| 191 |
async def main() -> None:
|
| 192 |
-
client = OpenAI(base_url=API_BASE_URL, api_key=
|
| 193 |
|
| 194 |
-
env = await CompilerEnv.from_docker_image(
|
| 195 |
|
| 196 |
history_messages: list = []
|
| 197 |
rewards: List[float] = []
|
|
@@ -233,7 +233,7 @@ async def main() -> None:
|
|
| 233 |
fn_name, fn_args, raw_tc = get_model_tool_call(client, history_messages)
|
| 234 |
|
| 235 |
if fn_name is None:
|
| 236 |
-
# No tool call
|
| 237 |
action_str = "compile_and_measure"
|
| 238 |
else:
|
| 239 |
action_str = tool_call_to_action(fn_name, fn_args)
|
|
|
|
| 45 |
|
| 46 |
from compiler_env import CompilerAction, CompilerEnv
|
| 47 |
|
| 48 |
+
LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
|
| 49 |
+
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 50 |
|
| 51 |
API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
|
| 52 |
MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"
|
|
|
|
| 189 |
|
| 190 |
|
| 191 |
async def main() -> None:
|
| 192 |
+
client = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN)
|
| 193 |
|
| 194 |
+
env = await CompilerEnv.from_docker_image(LOCAL_IMAGE_NAME)
|
| 195 |
|
| 196 |
history_messages: list = []
|
| 197 |
rewards: List[float] = []
|
|
|
|
| 233 |
fn_name, fn_args, raw_tc = get_model_tool_call(client, history_messages)
|
| 234 |
|
| 235 |
if fn_name is None:
|
| 236 |
+
# No tool call -- agent stopped; force compile_and_measure
|
| 237 |
action_str = "compile_and_measure"
|
| 238 |
else:
|
| 239 |
action_str = tool_call_to_action(fn_name, fn_args)
|
server/compiler_env_environment.py
CHANGED
|
@@ -144,19 +144,37 @@ class CompilerEnvironment(Environment):
|
|
| 144 |
raise ValueError(
|
| 145 |
f"Task ID '{task_id}' not found. Valid: {list(self.eval_programs.keys())}"
|
| 146 |
)
|
|
|
|
| 147 |
else:
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 160 |
|
| 161 |
return CompilerObservation(
|
| 162 |
data={
|
|
@@ -310,9 +328,9 @@ class CompilerEnvironment(Environment):
|
|
| 310 |
f"Sequence has {len(self.current_sequence)} passes.",
|
| 311 |
)
|
| 312 |
|
| 313 |
-
#
|
| 314 |
# Compilation & Measurement
|
| 315 |
-
#
|
| 316 |
|
| 317 |
def _compute_baseline(self, program_path, flag: str) -> float:
|
| 318 |
"""
|
|
@@ -326,7 +344,13 @@ class CompilerEnvironment(Environment):
|
|
| 326 |
"--gcc-toolchain=/usr",
|
| 327 |
f"-{flag}", program_path, "-o", exe_file, "-static",
|
| 328 |
]
|
| 329 |
-
subprocess.run(clang_cmd,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 330 |
|
| 331 |
start = time.perf_counter()
|
| 332 |
subprocess.run(
|
|
|
|
| 144 |
raise ValueError(
|
| 145 |
f"Task ID '{task_id}' not found. Valid: {list(self.eval_programs.keys())}"
|
| 146 |
)
|
| 147 |
+
programs_to_try = [self.current_program]
|
| 148 |
else:
|
| 149 |
+
programs_to_try = [random.choice(self.train_programs)]
|
| 150 |
+
# Build a retry list of up to 5 different programs
|
| 151 |
+
tried = {programs_to_try[0]}
|
| 152 |
+
while len(programs_to_try) < 5:
|
| 153 |
+
candidate = random.choice(self.train_programs)
|
| 154 |
+
if candidate not in tried:
|
| 155 |
+
programs_to_try.append(candidate)
|
| 156 |
+
tried.add(candidate)
|
| 157 |
+
|
| 158 |
+
# Try programs until one compiles successfully
|
| 159 |
+
last_error = None
|
| 160 |
+
for program in programs_to_try:
|
| 161 |
+
self.current_program = program
|
| 162 |
+
self.current_sequence = []
|
| 163 |
+
self.episode_reward = 0.0
|
| 164 |
+
self.done = False
|
| 165 |
+
try:
|
| 166 |
+
self.baseline_O0 = self._compute_baseline(self.current_program, "O0")
|
| 167 |
+
self.baseline_O1 = self._compute_baseline(self.current_program, "O1")
|
| 168 |
+
self.baseline_O2 = self._compute_baseline(self.current_program, "O2")
|
| 169 |
+
self.baseline_O3 = self._compute_baseline(self.current_program, "O3")
|
| 170 |
+
last_error = None
|
| 171 |
+
break
|
| 172 |
+
except (RuntimeError, subprocess.CalledProcessError, subprocess.TimeoutExpired) as e:
|
| 173 |
+
last_error = e
|
| 174 |
+
continue
|
| 175 |
+
|
| 176 |
+
if last_error is not None:
|
| 177 |
+
raise RuntimeError(f"All programs failed to compile: {last_error}")
|
| 178 |
|
| 179 |
return CompilerObservation(
|
| 180 |
data={
|
|
|
|
| 328 |
f"Sequence has {len(self.current_sequence)} passes.",
|
| 329 |
)
|
| 330 |
|
| 331 |
+
# ----------------------------------------------
|
| 332 |
# Compilation & Measurement
|
| 333 |
+
# ----------------------------------------------
|
| 334 |
|
| 335 |
def _compute_baseline(self, program_path, flag: str) -> float:
|
| 336 |
"""
|
|
|
|
| 344 |
"--gcc-toolchain=/usr",
|
| 345 |
f"-{flag}", program_path, "-o", exe_file, "-static",
|
| 346 |
]
|
| 347 |
+
result = subprocess.run(clang_cmd, capture_output=True, timeout=30)
|
| 348 |
+
if result.returncode != 0:
|
| 349 |
+
stderr = result.stderr.decode(errors="replace")
|
| 350 |
+
raise RuntimeError(
|
| 351 |
+
f"Baseline compilation failed for {os.path.basename(program_path)} "
|
| 352 |
+
f"with -{flag}: {stderr}"
|
| 353 |
+
)
|
| 354 |
|
| 355 |
start = time.perf_counter()
|
| 356 |
subprocess.run(
|