refactor(multimodal): tighten review comments, dedup SHM probe name

Trim the rationale comments from the prior two review commits down to the
non-obvious "why" only, and reuse next_mm_shm_name() for the /dev/shm
writability probe instead of duplicating the unix-nanos block.

Signed-off-by: Simo Lin <linsimo.mark@gmail.com>

Author: slin1237

grpc_servicer/smg_grpc_servicer/mm_shm.py

@@ -48,9 +48,8 @@ def tensor_payload_bytes_from_shm(shm_handle, shm_dir: str = DEFAULT_SHM_DIR) ->
     path = os.path.join(shm_dir, name)
     fd = None
     try:
-        # O_NOFOLLOW: /dev/shm is world-writable, so a same-host attacker could
-        # plant a symlink at the (validated) name pointing at an arbitrary file;
-        # refuse to follow it so a crafted handle can't read/unlink outside SHM.
+        # O_NOFOLLOW: /dev/shm is world-writable; refuse to follow a symlink planted
+        # at the validated name (would otherwise read/unlink an arbitrary file).
         fd = os.open(path, os.O_RDONLY | os.O_NOFOLLOW)
         raw = os.pread(fd, int(shm_handle.nbytes), int(shm_handle.offset))
     finally:

model_gateway/src/routers/grpc/common/stages/request_execution.rs

@@ -569,9 +569,8 @@ impl RequestExecutionStage {
         debug!("vLLM PD: prefill completed, sending decode request");
 
         // Decode reuses proto_request as-is; same request_id as the prefill leg is
-        // load-bearing for NIXL P/D correlation on vLLM < 0.13. Multimodal stays
-        // inline for PD (see resolve_mm_shm_enabled), so the decode leg can still
-        // recompute the prompt locally when KV relay is skipped (e.g. n>1 NIXL).
+        // load-bearing for NIXL P/D correlation on vLLM < 0.13. Multimodal is inline
+        // for PD, so decode keeps its inputs for the recompute path.
         let mut decode_request = proto_request;
         if let Some(rank) = workers.decode_worker().and_then(|w| w.dp_rank()) {
             decode_request.set_data_parallel_rank(rank as i32);

model_gateway/src/routers/grpc/multimodal.rs

@@ -1750,20 +1750,15 @@ fn resolve_mm_shm_enabled(
     transport: Option<&MmTransportConfig>,
 ) -> bool {
     log_mm_transport_config_once(transport);
-    // SHM is single-consumer (the servicer unlinks each segment after reading),
-    // but a PD request has two legs: prefill always reads, and decode re-reads
-    // whenever it recomputes the prompt instead of resuming from transferred KV.
-    // Force inline for disaggregated requests to avoid both the double-read and
-    // dropping pixels the decode leg still needs.
+    // SHM segments are single-consumer (unlinked after one read); a PD decode leg
+    // may re-read or recompute, so force inline for disaggregated requests.
     if matches!(workers, Some(WorkerSelection::Dual { .. })) {
         return false;
     }
     let global_mode = transport
         .map(|transport| transport.mode.as_str())
         .unwrap_or(DEFAULT_MM_TENSOR_TRANSPORT);
-    // A per-worker override wins, but only when it's a recognized mode; an
-    // invalid override is logged and ignored so a typo can't silently disable a
-    // valid global shm/auto.
+    // Ignore an unrecognized per-worker override so a typo can't disable a valid global.
     let mode = match worker_transport_mode_override(workers) {
         Some(mode) if matches!(mode.as_str(), "inline" | "shm" | "auto") => mode,
         Some(invalid) => {

model_gateway/src/routers/grpc/proto_wrapper.rs

@@ -503,14 +503,8 @@ fn write_mm_shm(data: &[u8]) -> std::io::Result<common::ShmHandle> {
 pub fn mm_shm_dev_writable() -> bool {
     static WRITABLE: OnceLock<bool> = OnceLock::new();
     *WRITABLE.get_or_init(|| {
-        // pid alone collides across separate PID namespaces that share /dev/shm
-        // (each can be PID 1); add a nanosecond stamp so the `create_new` probe
-        // doesn't spuriously fail and cache `false` for a writable /dev/shm.
-        let nanos = SystemTime::now()
-            .duration_since(UNIX_EPOCH)
-            .map(|duration| duration.as_nanos())
-            .unwrap_or_default();
-        let name = format!("{}probe-{}-{}", MM_SHM_NAME_PREFIX, process::id(), nanos);
+        // Unique name; pid alone collides across PID namespaces sharing /dev/shm.
+        let name = next_mm_shm_name();
         let path = mm_shm_path(&name);
         // `create_new` (no clobber) + owner-only mode: /dev/shm is world-writable,
         // so plain create(truncate) is open to symlink/clobber attacks and the