Spaces:
Running
Running
File size: 14,227 Bytes
7b4f5dd 43efb12 7b4f5dd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 | """
Performance Agent — GPU memory, latency and ROCm optimisation analyser.
Identifies ML-specific inefficiencies in code running on AMD MI300X.
"""
from __future__ import annotations
import json
import logging
import re
from typing import Any, AsyncGenerator, Dict, List, Optional
from openai import AsyncOpenAI
from api.models import PerformanceFinding, OptimizationType
from tools.code_parser import FileEntry, build_context_block
from tools.benchmark_tool import analyse_memory_optimisations
logger = logging.getLogger(__name__)
PERFORMANCE_SYSTEM_PROMPT = """You are CodeSentry Performance Agent — an AMD ROCm GPU performance engineer specialising in ML systems.
Analyse the provided code for performance issues specific to AI/ML workloads on AMD MI300X (192 GB HBM3).
## Check these categories (MANDATORY):
### GPU Memory Issues:
- Tensors allocated on GPU never moved back to CPU or deleted → VRAM leak
- Missing torch.cuda.empty_cache() / hip.device_synchronize() after batch inference
- Model loaded in float32 when float16/bfloat16 suffices → 2x VRAM waste
- Gradient tracking enabled during inference (missing @torch.no_grad or torch.inference_mode)
- KV cache not bounded → unbounded context growth
### Latency Issues:
- Model weights loaded inside per-request handler (should be singleton loaded at startup)
- Synchronous blocking calls inside async endpoints
- Tokenizer instantiated per-request instead of pre-loaded
- Missing torch.compile() for repeated inference patterns
### Throughput Issues:
- N+1 embedding calls: embed() called in a loop instead of batching all inputs
- Sequential agent calls that could be parallelised
- Missing continuous batching configuration in vLLM serving
- Single-worker serving when tensor parallelism is available
### ROCm/AMD-Specific:
- Using CUDA-only APIs not available on ROCm (use HIP equivalents)
- Missing HIP_VISIBLE_DEVICES environment configuration
- Not using Flash Attention 2 compatible with ROCm
- Memory bandwidth not maximised (FP8 quantisation available on MI300X)
## Output Format (STRICT JSON ARRAY):
[
{
"type": "gpu_memory|latency|throughput",
"title": "Short descriptive title",
"current_estimate": "Description of current resource usage",
"optimized_estimate": "Description after fix",
"saving_mb": <float MB saved or 0>,
"saving": "Human-readable saving description",
"suggestion": "Detailed explanation of the issue",
"code_fix": "Concrete code fix or snippet",
"line_number": <integer or null>,
"file_path": "<filename or null>"
}
]
Return ONLY the JSON array. If no issues found, return: []
"""
class PerformanceAgent:
def __init__(
self,
vllm_base_url: str = "http://localhost:8080/v1",
model: str = "Qwen/Qwen2.5-Coder-32B-Instruct",
api_key: str = "not-needed-local",
max_tokens: int = 3072,
temperature: float = 0.05,
) -> None:
self.model = model
self.max_tokens = max_tokens
self.temperature = temperature
self.client = AsyncOpenAI(
base_url=vllm_base_url,
api_key=api_key,
timeout=60.0,
max_retries=1,
)
# ─────────────────────────────────────────
# Static heuristic scan (no LLM)
# ─────────────────────────────────────────
def static_scan(self, files: List[FileEntry]) -> List[PerformanceFinding]:
"""Regex-based performance heuristics across all files."""
findings: List[PerformanceFinding] = []
for file_path, code in files:
heuristic_results = analyse_memory_optimisations(code)
for r in heuristic_results:
try:
opt_type = OptimizationType(r["type"])
except ValueError:
opt_type = OptimizationType.gpu_memory
findings.append(
PerformanceFinding(
type=opt_type,
title=f"[Static] {r['title']}",
current_estimate=r.get("current_estimate"),
optimized_estimate=r.get("optimized_estimate"),
saving_mb=r.get("saving_mb", 0.0),
saving=r.get("saving"),
description=r.get("suggestion", ""),
suggestion=r.get("code_fix", ""),
file=file_path,
)
)
# Additional per-file checks
findings.extend(self._check_model_loading_in_handler(code, file_path))
findings.extend(self._check_n_plus_one_loop(code, file_path))
findings.extend(self._check_fp32_usage(code, file_path))
return findings
def _check_model_loading_in_handler(self, code: str, file_path: str) -> List[PerformanceFinding]:
"""Detect model loading inside route/request handlers."""
results: List[PerformanceFinding] = []
# Find route decorators followed by from_pretrained within ~20 lines
lines = code.splitlines()
in_handler = False
handler_start = 0
for i, line in enumerate(lines):
stripped = line.strip()
if re.match(r"@(app|router)\.(get|post|put|delete|patch)", stripped):
in_handler = True
handler_start = i + 1
if in_handler and re.search(r"from_pretrained|AutoModel|AutoTokenizer", stripped):
if i - handler_start < 25:
results.append(
PerformanceFinding(
type=OptimizationType.latency,
title="[Static] Model loaded inside request handler",
current_estimate="Model weights loaded on every request (~10-30s cold start)",
optimized_estimate="Model singleton pre-loaded at startup (<1ms per request)",
saving_mb=0.0,
saving="Eliminates per-request load latency",
description="Model loaded once at startup using a global singleton or lifespan event.",
suggestion=(
"# At module level:\n"
"model = AutoModel.from_pretrained(...)\n\n"
"# In handler: use the pre-loaded `model`"
),
line=i + 1,
file=file_path,
)
)
in_handler = False
return results
def _check_n_plus_one_loop(self, code: str, file_path: str) -> List[PerformanceFinding]:
"""Detect embedding/encode calls inside for loops."""
results: List[PerformanceFinding] = []
lines = code.splitlines()
for i, line in enumerate(lines):
if re.match(r"\s*for\s+\w+\s+in\s+", line):
# Check next 5 lines for embed/encode calls
lookahead = "\n".join(lines[i + 1 : i + 6])
if re.search(r"\.(embed|encode|get_embedding)\(", lookahead):
results.append(
PerformanceFinding(
type=OptimizationType.throughput,
title="[Static] N+1 embedding calls in loop",
current_estimate="1 GPU kernel launch per item",
optimized_estimate="1 GPU kernel launch for all items",
saving_mb=0.0,
saving="Up to 50x throughput improvement",
description=(
"Embedding model called inside a loop. "
"Collect all inputs first, then batch-encode in one call."
),
suggestion=(
"# Instead of:\n"
"for text in texts:\n"
" emb = model.encode(text)\n\n"
"# Use:\n"
"embeddings = model.encode(texts, batch_size=32)"
),
line=i + 1,
file=file_path,
)
)
return results
def _check_fp32_usage(self, code: str, file_path: str) -> List[PerformanceFinding]:
"""Flag explicit float32 usage where bfloat16 would suffice."""
results: List[PerformanceFinding] = []
lines = code.splitlines()
for i, line in enumerate(lines):
if re.search(r"torch\.float32|torch_dtype\s*=\s*torch\.float32|\.float\(\)", line):
if not re.search(r"#.*noqa|#.*keep-fp32", line, re.IGNORECASE):
results.append(
PerformanceFinding(
type=OptimizationType.gpu_memory,
title="[Static] FP32 dtype — should use BF16",
current_estimate="4 bytes/param (float32)",
optimized_estimate="2 bytes/param (bfloat16) — 50% VRAM saving",
saving_mb=None,
saving="~50% VRAM reduction on MI300X",
description="AMD MI300X supports bfloat16 natively with no accuracy loss for inference.",
suggestion=(
"# Replace:\n"
"model = model.float()\n"
"# With:\n"
"model = model.to(torch.bfloat16) # or torch_dtype=torch.bfloat16"
),
line=i + 1,
file=file_path,
)
)
return results
# ─────────────────────────────────────────
# LLM analysis
# ─────────────────────────────────────────
async def llm_scan(self, code_context: str) -> List[PerformanceFinding]:
"""Deep LLM-based performance analysis."""
user_message = (
"Analyse the following codebase for GPU memory, latency, and throughput issues "
"on AMD MI300X hardware:\n\n"
f"```\n{code_context}\n```\n\n"
"Return ONLY the JSON array of performance findings."
)
try:
response = await self.client.chat.completions.create(
model=self.model,
messages=[
{"role": "system", "content": PERFORMANCE_SYSTEM_PROMPT},
{"role": "user", "content": user_message},
],
max_tokens=self.max_tokens,
temperature=self.temperature,
)
raw = response.choices[0].message.content or "[]"
return self._parse_llm_response(raw)
except Exception as exc:
logger.error("[PerformanceAgent] LLM call failed: %s", exc)
return []
async def analyze(
self,
files: List[FileEntry],
code_context: str,
use_llm: bool = True,
) -> List[PerformanceFinding]:
"""Full pipeline: static heuristics + LLM deep analysis."""
static = self.static_scan(files)
logger.info("[PerformanceAgent] Static scan: %d findings", len(static))
if not use_llm:
return static
llm_findings = await self.llm_scan(code_context)
logger.info("[PerformanceAgent] LLM scan: %d findings", len(llm_findings))
# Merge: deduplicate by title
llm_titles = {f.title for f in llm_findings}
merged = list(llm_findings)
for f in static:
clean_title = f.title.replace("[Static] ", "")
if clean_title not in llm_titles:
merged.append(f)
return merged
# ─────────────────────────────────────────
# Helpers
# ─────────────────────────────────────────
def _parse_llm_response(self, raw: str) -> List[PerformanceFinding]:
raw = re.sub(r"```(?:json)?\s*", "", raw).strip().rstrip("`").strip()
start, end = raw.find("["), raw.rfind("]") + 1
if start == -1 or end == 0:
return []
try:
data: List[Dict] = json.loads(raw[start:end])
except json.JSONDecodeError:
return []
findings: List[PerformanceFinding] = []
for item in data:
try:
opt_type_str = item.get("type", "gpu_memory")
try:
opt_type = OptimizationType(opt_type_str)
except ValueError:
opt_type = OptimizationType.gpu_memory
findings.append(
PerformanceFinding(
type=opt_type,
title=item.get("title", "Unknown"),
current_estimate=item.get("current_estimate"),
optimized_estimate=item.get("optimized_estimate"),
saving_mb=item.get("saving_mb"),
saving=item.get("saving"),
description=item.get("suggestion", ""),
suggestion=item.get("code_fix"),
line=item.get("line_number"),
file=item.get("file_path"),
code=item.get("code_snippet"),
)
)
except Exception as e:
logger.debug("[PerformanceAgent] Skipping malformed finding: %s", e)
return findings
|