Introducing Sim mode

sendnn_inference/envs.py

@@ -180,6 +180,21 @@ def clear_env_cache():
     "SENDNN_INFERENCE_TP_MM_SHARING": lambda: bool(
         int(os.getenv("SENDNN_INFERENCE_TP_MM_SHARING", "1"))
     ),
+    # Sim-mode: replace the real Spyre forward with a no-op model and substitute
+    # virtual durations into request_metrics.jsonl. Lets us exercise scheduler /
+    # runner / batching logic on a laptop without AIU hardware. Use with
+    # DYNAMO_BACKEND=eager.
+    "SENDNN_INFERENCE_SIM_MODE": lambda: bool(int(os.getenv("SENDNN_INFERENCE_SIM_MODE", "0"))),
+    # Virtual duration (ms) charged for each prefill forward step in sim mode.
+    # Operator-supplied; default 0 means no virtual time accumulates.
+    "SENDNN_INFERENCE_SIM_PREFILL_MS": lambda: float(
+        os.getenv("SENDNN_INFERENCE_SIM_PREFILL_MS", "0")
+    ),
+    # Virtual duration (ms) charged for each decode forward step in sim mode.
+    # Operator-supplied; default 0 means no virtual time accumulates.
+    "SENDNN_INFERENCE_SIM_DECODE_MS": lambda: float(
+        os.getenv("SENDNN_INFERENCE_SIM_DECODE_MS", "0")
+    ),
 }
 # --8<-- [end:env-vars-definition]
 

sendnn_inference/platform.py

@@ -242,6 +242,18 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         if not is_decoder and not is_pooling:
             raise ValueError("Only the 'generate' and 'pooling' runners are supported")
 
+        if envs_spyre.SENDNN_INFERENCE_SIM_MODE:
+            if parallel_config.tensor_parallel_size != 1:
+                raise ValueError(
+                    "SENDNN_INFERENCE_SIM_MODE only supports tensor_parallel_size=1, "
+                    f"got {parallel_config.tensor_parallel_size}."
+                )
+            if envs_spyre.SENDNN_INFERENCE_DYNAMO_BACKEND != "eager":
+                raise ValueError(
+                    "SENDNN_INFERENCE_SIM_MODE requires SENDNN_INFERENCE_DYNAMO_BACKEND=eager, "
+                    f"got '{envs_spyre.SENDNN_INFERENCE_DYNAMO_BACKEND}'."
+                )
+
         if parallel_config.worker_cls == "auto":
             parallel_config.worker_cls = "sendnn_inference.v1.worker.spyre_worker.SpyreWorker"
 
@@ -260,9 +272,14 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
             os.environ["FLEX_DEVICE"] = "COMPILE"
 
         if is_decoder:
-            scheduler_config.scheduler_cls = (
-                "sendnn_inference.v1.core.scheduler.ChunkedPrefillSpyreScheduler"
-            )
+            if envs_spyre.SENDNN_INFERENCE_SIM_MODE:
+                scheduler_config.scheduler_cls = (
+                    "sendnn_inference.v1.sim.SimulatedChunkedPrefillSpyreScheduler"
+                )
+            else:
+                scheduler_config.scheduler_cls = (
+                    "sendnn_inference.v1.core.scheduler.ChunkedPrefillSpyreScheduler"
+                )
 
             if (
                 vllm_config.model_config.quantization

sendnn_inference/v1/metrics/stats_logger.py

@@ -121,6 +121,12 @@ def record(
             self.iso_format
         )[:-3]
 
+        if envs_spyre.SENDNN_INFERENCE_SIM_MODE:
+            # In sim mode, virtual timings are emitted directly from the engine
+            # process to sim_metrics.jsonl. The wall-clock fields here are ~0
+            # and would be misleading; skip writing.
+            return
+
         records_to_write: list[str] = []
         for r in iteration_stats.finished_requests:
             # Calculate some estimates to add to the engine stats