chunked prefill

configs/qwen3omni.yaml

@@ -1,5 +1,15 @@
 model: "qwen3_omni"
 max_seq_len: 32768
+# Engine: chunked prefill. Splits long prefills into 512-token chunks.
+# Set to null (or remove) to disable. Only applies to qwen3_omni Thinker
+# (the LLM submodule) — other submodules opt in individually.
+max_prefill_chunk_size: 512
+# Phase 2: scheduler-driven chunked prefill. When true, the MicroScheduler
+# packs mixed batches (decodes + prefill chunks across requests) up to
+# max_step_tokens. When false (default), the engine handles single-request
+# chunking internally (Phase 1).
+scheduler_owns_chunking: false
+max_step_tokens: 2048
 node_groups:
   - node_names: [audio_encoder, vision_encoder, Code2Wav]
     ranks: [0]

mminf/conductor/request_info.py

@@ -77,6 +77,18 @@ class CurrentForwardPassInfo:
     loop_stop_times: dict[str, IterIndexTree] = field(default_factory=dict)
     dynamic_loop_iter_counts: dict[str, int] = field(default_factory=dict)
 
+    # chunked prefill progress.
+    # Set at request admission; advanced by the MicroScheduler each step
+    # as chunks complete. Derived `is_prefill_complete` gates the
+    # prefill→decode transition. Default values (0, 0) mean a request not
+    # in chunked-prefill mode.
+    prefill_tokens_total: int = 0
+    prefill_tokens_consumed: int = 0
+
+    @property
+    def is_prefill_complete(self) -> bool:
+        return self.prefill_tokens_consumed >= self.prefill_tokens_total
+
     def register_loop_stop(self, loop_name: str):
         self.dynamic_loop_stop_signals.add(loop_name)