Add files using upload-large-folder tool

code/models/common/models/qwen3_32b/hf_adaptor.py

@@ -172,7 +172,14 @@ def _qwen_stop_token_ids(tokenizer) -> tuple[int, ...]:
 def _trace_seq_lens(num_devices: int, max_prefill_chunk_size: int, max_seq_len: int) -> tuple[int, ...]:
     if num_devices not in (4, 8):
         raise ValueError(f"Qwen3-32B supports T3K (8 devices) or P150x4 (4 devices), got {num_devices}")
-    candidates = (128, 1024)
+    # Every prefill shape the planner can emit: the 128/1024 padding buckets plus
+    # the power-of-two chunk sizes bounded by max_prefill_chunk_size (2048, 4096).
+    # A shape absent here is neither trace-eligible nor warmed, so under
+    # trace_mode="all" the first such prompt dies with TraceCoverageError (and
+    # under decode_only with "not compiled after trace activation"). See the
+    # planner: _padded_prefill_length rounds to 128/1024/next-pow2 and
+    # _max_prefill_chunk_size caps chunks at max_prefill_chunk_size.
+    candidates = (128, 1024, 2048, 4096)
     return tuple(length for length in candidates if length <= min(max_prefill_chunk_size, max_seq_len))