fix(pr13): address CodeRabbit review findings

Clarit-AI · claude · Clarit-AI · commit be1c95536aa0 · 2026-03-31T22:22:18.000Z
- Remove accidental .remember/tmp/save-session.pid from tracking
- Add .remember/tmp/ to .gitignore
- Fix undefined variables in bench_multiturn.py barrier mode
- Add missing self.lora_path in WorkloadGenerator.__init__
- Zero padded tail in nemotron_h piecewise graph mode

Co-Authored-By: Claude Opus 4.6 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/.gitignore b/.gitignore
@@ -275,5 +275,8 @@ sgl-kernel/csrc/**/*_musa/
 # VM migration context (local only, not for commit)
 docs/migration-prep/
 
+# Runtime state files
+.remember/tmp/
+
 # Beads / Dolt files (added by bd init)
 *.db
diff --git a/.remember/tmp/save-session.pid b/.remember/tmp/save-session.pid
diff --git a/benchmark/hicache/bench_multiturn.py b/benchmark/hicache/bench_multiturn.py
@@ -260,6 +260,7 @@ def __init__(self, args):
                 }
         self.num_clients = args.num_clients
 
+        self.lora_path = args.lora_path
         self.num_rounds = args.num_rounds
         self.max_parallel = args.max_parallel
         self.output_length = args.output_length
@@ -309,6 +310,8 @@ async def request_loop():
 
     def response_handler(self):
         next_round_reqs = []
+        current_barrier_round = 0
+        barrier_round_completed = 0
         while True:
             try:
                 client_id, response = self.response_queue.get(
@@ -352,7 +355,7 @@ def response_handler(self):
                         gen_payload(
                             self.client_records[client_id]["history"],
                             self.output_length,
-                            args.lora_path,
+                            self.lora_path,
                         ),
                     )
                     if self.enable_round_barrier:
@@ -368,10 +371,10 @@ def response_handler(self):
                 # current barrier round have completed
                 if (
                     self.enable_round_barrier
-                    and current_barrier_round < self.max_rounds
+                    and current_barrier_round < self.num_rounds
                 ):
                     barrier_round_completed += 1
-                    expected = self.clients_per_round[current_barrier_round]
+                    expected = self.num_clients
                     if barrier_round_completed == expected:
                         print(
                             f"\n  Barrier: round {current_barrier_round} complete "
diff --git a/python/sglang/srt/models/nemotron_h.py b/python/sglang/srt/models/nemotron_h.py
@@ -898,4 +898,6 @@ def nemotron_mamba2_with_output(
 
     # Copy result back; output may be larger (padded) so only fill actual tokens
     output[:num_actual_tokens].view(ret.shape).copy_(ret)
+    if output.shape[0] > num_actual_tokens:
+        output[num_actual_tokens:].zero_()
     return
