chore: cleanup garbage code (#29770)

scripts/ci/amd/amd_ci_install_dependency.sh

@@ -150,8 +150,6 @@ else
   docker cp human-eval ci_sglang:/
   install_with_retry docker exec -w /human-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
 
-  docker exec -w / ci_sglang mkdir -p /dummy-grok
-  # Create dummy grok config inline (bypasses Azure blob storage which may have auth issues)
   mkdir -p dummy-grok
   cat > dummy-grok/config.json << 'EOF'
   {
@@ -176,9 +174,8 @@ else
     "torch_dtype": "bfloat16"
   }
 EOF
-  # docker exec -w / ci_sglang mkdir -p /dummy-grok
-  # mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
-  # docker cp ./dummy-grok ci_sglang:/
+  docker exec -w / ci_sglang mkdir -p /dummy-grok
+  docker cp ./dummy-grok/config.json ci_sglang:/dummy-grok/config.json
 
   docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache huggingface_hub[hf_xet]
   docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache pytest

scripts/ci/slurm/analyze_logs_with_modal.py

@@ -143,7 +143,7 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-def build_sandbox_image() -> "modal.Image":
+def build_sandbox_image() -> modal.Image:
     if modal is None:
         raise RuntimeError(
             "The 'modal' package is required. Run this script with "
@@ -213,7 +213,7 @@ Do not skip issue filing. The whole point of this system is automated triage.
 """
 
 
-def upload_tree(sandbox: "modal.Sandbox", log_dir: Path) -> None:
+def upload_tree(sandbox: modal.Sandbox, log_dir: Path) -> None:
     log_files = [path for path in log_dir.rglob("*") if path.is_file()]
     logger.info("Uploading %d log files into the sandbox", len(log_files))
     for index, log_file in enumerate(log_files, start=1):
@@ -225,7 +225,7 @@ def upload_tree(sandbox: "modal.Sandbox", log_dir: Path) -> None:
             logger.info("Uploaded %d/%d files", index, len(log_files))
 
 
-def clone_context_repos(sandbox: "modal.Sandbox", repo_urls: list[str]) -> None:
+def clone_context_repos(sandbox: modal.Sandbox, repo_urls: list[str]) -> None:
     if not repo_urls:
         return
 
@@ -244,7 +244,7 @@ def clone_context_repos(sandbox: "modal.Sandbox", repo_urls: list[str]) -> None:
         ).wait()
 
 
-def read_optional_file(sandbox: "modal.Sandbox", path: str) -> str | None:
+def read_optional_file(sandbox: modal.Sandbox, path: str) -> str | None:
     try:
         return sandbox.filesystem.read_text(path)
     except Exception:

scripts/ci/slurm/launch_mi355x.sh

@@ -231,10 +231,10 @@ docker run $DOCKER_COMMON --name mi355x_decode \
   $COMMON_FLAGS --disaggregation-mode decode --disaggregation-bootstrap-port $DBOOT
 EOF
 
-# Smoke-test payload + validator (separate files to avoid quoting inside the
+# Probe payload + validator (separate files to avoid quoting inside the
 # bench.sh `bash -lc '...'` block). One real request exercises the full
 # prefill->decode KV handoff before we commit to the whole sweep.
-cat > "$WORKDIR/smoke.json" <<'JSON'
+cat > "$WORKDIR/probe.json" <<'JSON'
 {"text": "The capital of France is", "sampling_params": {"max_new_tokens": 16, "temperature": 0.0}}
 JSON
 cat > "$WORKDIR/assert_nonempty.py" <<'PY'
@@ -242,9 +242,9 @@ import sys, json
 d = json.load(sys.stdin)
 t = d.get("text", "") if isinstance(d, dict) else ""
 if not (t and t.strip()):
-    print("[smoke] empty/invalid output:", str(d)[:200])
+    print("[probe] empty/invalid output:", str(d)[:200])
     sys.exit(1)
-print("[smoke] ok:", t[:80].replace("\n", " "))
+print("[probe] ok:", t[:80].replace("\n", " "))
 PY
 
 # Bench script runs on the prefill node; \$PIP/\$DIP injected at srun time.
@@ -267,12 +267,12 @@ docker run $DOCKER_COMMON --name mi355x_bench \
       --disable-circuit-breaker &
     for i in \$(seq 1 30); do curl -sf http://127.0.0.1:$LBPORT/health >/dev/null && break; sleep 2; done
     CIDIR=/host_home/.mi355x_ci/${MATRIX_CONFIG_NAME}
-    echo "[smoke] PD end-to-end check via LB"
+    echo "[probe] PD end-to-end check via LB"
     curl -sf -X POST http://127.0.0.1:$LBPORT/generate \
-      -H "content-type: application/json" -d @\$CIDIR/smoke.json > \$CIDIR/smoke_out.json \
-      || { echo "[smoke] request failed -- PD path not serving; aborting before sweep"; exit 1; }
-    python3 \$CIDIR/assert_nonempty.py < \$CIDIR/smoke_out.json \
-      || { echo "[smoke] empty/invalid generation; aborting before sweep"; exit 1; }
+      -H "content-type: application/json" -d @\$CIDIR/probe.json > \$CIDIR/probe_out.json \
+      || { echo "[probe] request failed -- PD path not serving; aborting before sweep"; exit 1; }
+    python3 \$CIDIR/assert_nonempty.py < \$CIDIR/probe_out.json \
+      || { echo "[probe] empty/invalid generation; aborting before sweep"; exit 1; }
     # Correctness gate runs BEFORE the perf sweep: if the model is wrong there
     # is no point spending ~15min measuring how fast it is wrong, so a failure
     # here exits immediately and the sweep never runs.

scripts/ci/utils/compute_partitions.py

@@ -35,7 +35,7 @@ HWBackend = _ci_register.HWBackend
 # pr-test-amd.yml / pr-test-npu.yml have their own dispatch.
 _TARGET_BACKENDS = {HWBackend.CUDA, HWBackend.CPU}
 
-# base-a is the critical-path entry gate; pin its fanout to smoke-coverage
+# base-a is the critical-path entry gate; pin its fanout to sanity-coverage
 # defaults instead of est_time. max_parallel = size (no throttle).
 _BASE_A_OVERRIDES = {
     "base-a-test-cpu": 8,