review-pr

ATOM PR Review

ATOM is a ROCm/AMD GPU kernel optimization layer (MI300X/MI355X) that:

• Consumes aiter ops (attention, MoE, GEMM, norm, quant)
• Integrates with vLLM and SGLang as a plugin/backend
• Provides custom MLA, sparse attention, TBO, and quantization fusion

A change here can break inference for all models using the affected kernel path.

Step 1 — Fetch

PR=$1
REPO="ROCm/ATOM"

gh pr view $PR --repo $REPO --json title,body,number,labels,files,author,reviews,comments > /tmp/pr_meta.json
gh pr diff $PR --repo $REPO > /tmp/pr.diff

# Linked issue
ISSUE=$(cat /tmp/pr_meta.json | python3 -c "
import json,re,sys
body = json.load(sys.stdin).get('body','') or ''
m = re.search(r'(?:fix|close|resolve)[s]?[: ]*#(\d+)', body, re.I)
print(m.group(1) if m else '')
")
[ -n "$ISSUE" ] && gh issue view $ISSUE --repo $REPO --json title,body > /tmp/pr_issue.json

# Prior reviewer comments (top-level)
cat /tmp/pr_meta.json | python3 -c "
import json,sys
d = json.load(sys.stdin)
for r in d.get('reviews',[]):
    b = (r.get('body','') or '').strip()
    if b: print(f'[REVIEW {r[\"author\"][\"login\"]}] {b[:200]}')
for c in d.get('comments',[]):
    b = (c.get('body','') or '').strip()
    if b: print(f'[COMMENT {c[\"author\"][\"login\"]}] {b[:200]}')
"

# Inline review comments (line-level — often more specific than top-level)
gh api "repos/$REPO/pulls/$PR/comments" | python3 -c "
import json,sys
comments = json.load(sys.stdin)
for c in comments:
    author = c.get('user',{}).get('login','')
    body = (c.get('body','') or '').strip()
    path = c.get('path','')
    line = c.get('line') or c.get('original_line','')
    if body and 'copilot' not in author.lower() and 'bot' not in author.lower():
        print(f'[INLINE {author}] {path}:{line}')
        print(f'  {body[:250]}')
" 2>/dev/null

Read the diff and PR body before proceeding.

Cross-file verification — before reporting any kernel/dispatch finding. The diff shows changed lines, not the whole story. Grep the entire symbol family (.cu + .cuh + .h, or the whole module), not just files in the diff — sync/fence/atomics or the "other half" of a scatter often live in a header, and dispatch/else-branch completeness must be read in the full function, not the hunk. A "no synchronization" or "missing branch" finding based only on the diff is how false positives happen.