[CI] Parity: fold CUDA distributed-test split (pytorch#189232) into distributed config

.automation_scripts/pytorch-unit-test-scripts/auto_classify_skip_reasons.py

@@ -46,6 +46,15 @@
     # TIER 1: High-specificity combined rules (message + file/class)
     # ==================================================================
 
+    # --- PT2.0 - Convolution: conv2d backward parametrized skipped on ROCm ---
+    # skipIfRocmArch(...) skips test_conv2d_backward_parametrized (see upstream
+    # #188671, CI timeout). Categorize it as a convolution issue rather than the
+    # generic Misc "test skipped on ('gfx...')" bucket below. Must precede that
+    # Misc arch rule.
+    {"reason": "PT2.0 - Convolution",
+     "msg": r"test skipped on \('gfx",
+     "name": r"(?i)conv2d_backward"},
+
     # --- bfloat16_SDPA_ME: dropout mask in test_transformers with bfloat16 in TEST NAME ---
     # Must be before generic SDPA_ME rule
     {"reason": "bfloat16_SDPA_ME",

.automation_scripts/pytorch-unit-test-scripts/detect_log_failures.py

@@ -69,13 +69,23 @@
 
 
 def classify_log_file(filename):
-    """Return (platform, test_config, shard_num) from a log filename like rocm3.txt."""
+    """Return (platform, test_config, shard_num) from a log filename like rocm3.txt.
+
+    Commit-vs-commit parity prefixes log files with the short commit SHA
+    (for example, 09e0c59b_rocm3.txt). In that mode the SHA label is the
+    platform name used by generate_summary.py, so preserve it here.
+    """
     stem = Path(filename).stem
+    label = None
+    m = re.match(r"(?P<label>[0-9a-f]{8,40})_(?P<stem>.+)", stem)
+    if m:
+        label = m.group("label")[:8]
+        stem = m.group("stem")
     for prefix, (platform, test_config) in sorted(LOG_FILE_MAP.items(), key=lambda x: -len(x[0])):
         if stem.startswith(prefix):
             remainder = stem[len(prefix):]
             if remainder.isdigit():
-                return platform, test_config, int(remainder)
+                return label or platform, test_config, int(remainder)
     return None, None, None
 
 

.automation_scripts/pytorch-unit-test-scripts/download_testlogs

@@ -185,6 +185,36 @@ def get_workflow_jobs(wf, all_attempts=False):
             jobs += page_json["jobs"]
     return jobs
 
+def derive_shard_count(wf, job_prefix, test_config, fallback):
+    """Return the shard total upstream actually used for this (job_prefix,
+    test_config), parsed from job names like
+    '<job_prefix> / test (<config>, <idx>, <total>, ...)' -> <total>.
+
+    Shard counts differ between an arch's primary workflow and its fallback
+    (e.g. mi200 default is 6 shards in rocm-mi200 but 10 in the
+    trunk-rocm-sandbox fallback), so we read the real total from the resolved
+    run instead of assuming the config value - otherwise the constructed
+    "(default, i, 6)" keys miss the actual "(default, i, 10)" jobs. Falls back
+    to `fallback` when the run has no matching jobs.
+    """
+    try:
+        jobs = get_workflow_jobs(wf)
+    except Exception:
+        return fallback
+    pat = re.compile(
+        re.escape(job_prefix) + r" / \S+ \(" + re.escape(test_config) + r", \d+, (\d+)"
+    )
+    totals = {}
+    for job in jobs:
+        m = pat.search(job.get("name", ""))
+        if m:
+            total = int(m.group(1))
+            totals[total] = totals.get(total, 0) + 1
+    if not totals:
+        return fallback
+    # Most common total wins (guards against a stray reshaped/duplicate job).
+    return sorted(totals.items(), key=lambda kv: (-kv[1], -kv[0]))[0][0]
+
 def get_check_runs_for_commit(sha, prefix):
     """Get check runs for a commit filtered by name prefix.
 
@@ -526,7 +556,7 @@ def parse_args():
     parser.add_argument('--no_rocm', action='store_true')
     parser.add_argument('--no_cuda', action='store_true')
     parser.add_argument('--pr_id', type=int, help='The pull request ID')
-    parser.add_argument('--arch', type=str, choices=['mi200', 'mi300', 'mi350', 'navi31', 'nightly'], default='mi350', help='ROCm GPU architecture (mi200, mi300, mi350, navi31, or nightly, default: mi350)')
+    parser.add_argument('--arch', type=str, choices=['mi200', 'mi300', 'mi350', 'navi31', 'nightly', 'preview'], default='mi350', help='ROCm GPU architecture (mi200, mi300, mi350, navi31, nightly, or preview, default: mi350)')
     parser.add_argument('--include_inductor_periodic', action='store_true', help='Also download inductor-periodic benchmark artifacts (into a separate directory, not included in parity CSV)')
     parser.add_argument('--baseline_sha', type=str, help='Baseline commit SHA to compare against. Downloads the same ROCm workflows for this commit into baseline_xml/.')
     return parser.parse_args()
@@ -654,17 +684,37 @@ def main():
     token = os.getenv('GITHUB_TOKEN', '...')
     global authentication_headers
     authentication_headers = {'Authorization': f'token {token}'}
-    if (args.pr_id and args.sha1) or (not args.pr_id and not args.sha1):
-        error_msg = "Error: Please provide either pr_id or sha!"
+    status = "success"
+    # An empty --sha1 (or the literal "latest") means "use the latest green
+    # run on main" rather than a specific commit. download_workflow_run()
+    # already resolves that when given no head_sha, so treat these as unset.
+    sha_is_latest = (not args.sha1) or (str(args.sha1).strip().lower() == "latest")
+    if args.pr_id and not sha_is_latest:
+        error_msg = "Error: Please provide either pr_id or sha (not both)!"
         print(error_msg)
         sys.exit(1)
     if args.pr_id:
         pr_id = args.pr_id
-        sha = get_latest_commit_sha(pr_id, token)        
-    else:
+        sha = get_latest_commit_sha(pr_id, token)
+    elif not sha_is_latest:
         sha = args.sha1
         pr_id = None
-    status = "success"
+    else:
+        # No pr_id and no explicit sha: resolve the latest green ROCm run on
+        # main and use its head commit as the target sha.
+        pr_id = None
+        print("No sha or pr_id given; resolving latest green ROCm run on main...")
+        latest_wf = download_workflow_run(
+            created=args.created,
+            max_pages=args.max_pages,
+            workflow=ROCmWorkflowNames["default"],
+            sha=None,
+            ignore_status=args.ignore_status,
+            status=status,
+            error_msg="Error: could not find a latest green ROCm run on main",
+        )
+        sha = latest_wf["head_sha"]
+        print(f"Resolved 'latest' to sha: {sha}")
     print(f"sha: {sha}")
 
     # When comparing two commits, prefix log filenames with short SHAs
@@ -742,10 +792,7 @@ def main():
         # HUD link: https://hud.pytorch.org/hud/pytorch/pytorch/main/1?per_page=50&name_filter=rocm
         # Make sure "Hide unstable jobs" is unselected, in case ROCm jobs are marked as unstable
 
-        if arch == "mi350":
-            dist_shards = 3 if not periodic_fallback_used else rocm_shards["distributed"]
-        else:
-            dist_shards = rocm_shards["distributed"]
+        dist_shards = derive_shard_count(periodic_wf, dist_job_prefix, "distributed", rocm_shards["distributed"])
         print(f"Using final ROCm shard count {dist_shards} for distributed")
 
         if not args.artifacts_only:
@@ -805,10 +852,7 @@ def main():
         # Download logs
         # If logs aren't found you might want to check the HUD for the correct tags
         # HUD link: https://hud.pytorch.org/hud/pytorch/pytorch/main/1?per_page=50&name_filter=rocm
-        if arch == "mi350":
-            default_shards = 6 if default_fallback_used else rocm_shards["default"]
-        else:
-            default_shards = rocm_shards["default"]
+        default_shards = derive_shard_count(rocm_wf, rocm_job_prefix['default'], "default", rocm_shards["default"])
         print(f"Using final ROCm shard count {default_shards} for default")
 
         if not args.artifacts_only:
@@ -863,12 +907,12 @@ def main():
 
             folder_list = get_or_create_test_folder(inductor_wf_rocm)
 
-            inductor_shards = rocm_shards["inductor"]
-            print(f"Using final ROCm shard count {inductor_shards} for inductor")
             if inductor_fallback_used and arch in inductor_fallbacks:
                 inductor_job_prefix = inductor_fallbacks[arch][1]
             else:
                 inductor_job_prefix = rocm_job_prefix['inductor']
+            inductor_shards = derive_shard_count(inductor_wf_rocm, inductor_job_prefix, "inductor", rocm_shards["inductor"])
+            print(f"Using final ROCm shard count {inductor_shards} for inductor")
 
             # Download logs
             if not args.artifacts_only:

.automation_scripts/pytorch-unit-test-scripts/generate_summary.py

@@ -114,10 +114,10 @@ def test_config_stats_keys(s1_name, s2_name, has_set2=True):
             f'TOTAL {s1}',
         ]
     return [
-        f'SKIPPED (on {s1_name}, but not on {s2_name})',
+        f'SKIPPED (on {s1_name}, PASSED on {s2_name})',
         f'SKIPPED (on {s1_name})',
         f'SKIPPED (on {s2_name})',
-        f'MISSED (MISSED on {s1_name}, NOT SKIPPED on {s2_name})',
+        f'MISSED (on {s1_name}, PASSED on {s2_name})',
         f'{s1}ONLY (PASSED on {s1}, NOT PASSED on {s2})',
         s2,
         s1,
@@ -141,15 +141,18 @@ def compute_test_config_stats(rows, s1_col, s2_col, s1_name, s2_name, has_set2=T
         vals[keys[4]] = sum(1 for r in rows if r[s1_col].strip())
         return vals
 
+    # Count a ROCm SKIPPED/MISSED test as a disagreement only when CUDA actually
+    # PASSED it. This excludes parametrization variants CUDA never runs (CUDA
+    # MISSED) or also skips, which aren't real ROCm-vs-CUDA coverage gaps.
     s1_skip_not_s2 = sum(
         1 for r in rows
-        if r[s1_col] == 'SKIPPED' and r[s2_col] != 'SKIPPED'
+        if r[s1_col] == 'SKIPPED' and r[s2_col] == 'PASSED'
     )
     s1_skip = sum(1 for r in rows if r[s1_col] == 'SKIPPED')
     s2_skip = sum(1 for r in rows if r[s2_col] == 'SKIPPED')
     s1_miss_not_s2_skip = sum(
         1 for r in rows
-        if r[s1_col] == 'MISSED' and r[s2_col] != 'SKIPPED'
+        if r[s1_col] == 'MISSED' and r[s2_col] == 'PASSED'
     )
     only_s1 = sum(
         1 for r in rows
@@ -233,11 +236,11 @@ def safe_float(v):
         wf_rows = [r for r in rows if r['test_config'] == wf]
         s1_skip_not_s2 = sum(
             1 for r in wf_rows
-            if r[s1_col] == 'SKIPPED' and r[s2_col] != 'SKIPPED'
+            if r[s1_col] == 'SKIPPED' and r[s2_col] == 'PASSED'
         )
         s1_miss_not_s2_skip = sum(
             1 for r in wf_rows
-            if r[s1_col] == 'MISSED' and r[s2_col] != 'SKIPPED'
+            if r[s1_col] == 'MISSED' and r[s2_col] == 'PASSED'
         )
         total_disagree += s1_skip_not_s2 + s1_miss_not_s2_skip
         total_s2 += sum(1 for r in wf_rows if r[s2_col].strip() and r[s2_col].strip() != 'MISSED')
@@ -639,6 +642,18 @@ def build_rows(args, archs, arch_data):
 
     if args.sha:
         out.append(('__header__', f'Commit SHA: {args.sha}'))
+        # Link straight to the upstream HUD page filtered (regex) to the trunk
+        # CUDA / inductor / rocm test jobs this report is built from, so a
+        # reviewer can jump from the summary to the matching CI jobs. Parens and
+        # pipes are percent-encoded to keep the markdown link valid.
+        hud_url = (
+            f'https://hud.pytorch.org/hud/pytorch/pytorch/{args.sha}/1'
+            '?per_page=50'
+            '&name_filter=%28trunk.*cuda%7Cinductor%7Crocm%29.*test.*'
+            '%28default%7Cdistributed%7Cinductor%29%2C'
+            '&useRegexFilter=true'
+        )
+        out.append(('__header__', f'HUD: [parity jobs for this commit]({hud_url})'))
     if args.pr_id:
         out.append(('__header__', f'PR ID: {args.pr_id}'))
 

.automation_scripts/pytorch-unit-test-scripts/parity_job_config.json

@@ -70,6 +70,13 @@
       "inductor": [{ "workflow": "rocm-nightly", "job_prefix": "linux-noble-rocm-nightly-py3.12-gfx942" }],
       "shard_counts": { "default": 6, "distributed": 3, "inductor": 2 },
       "checkrun_regex": "rocm-nightly.*/ test [(](default|distributed|inductor),"
+    },
+    "preview": {
+      "default": [{ "workflow": "rocm-preview", "job_prefix": "linux-noble-rocm-preview-py3.12-gfx942" }],
+      "distributed": [{ "workflow": "rocm-preview", "job_prefix": "linux-noble-rocm-preview-py3.12-gfx942" }],
+      "inductor": [{ "workflow": "rocm-preview", "job_prefix": "linux-noble-rocm-preview-py3.12-gfx942" }],
+      "shard_counts": { "default": 6, "distributed": 3, "inductor": 2 },
+      "checkrun_regex": "linux-noble-rocm-preview-py3.12-gfx942 / test [(](default|distributed|inductor),"
     }
   }
 }