🐛 use eager compile (#488)

# Description

In some environments the default of `inductor` here is breaking vllm
where a compiler isn't available.

This PR makes the simple compile backend configurable, and defaults to
`eager` for portability.

Signed-off-by: Joe Runde <[email protected]>

Author: joerunde

vllm_spyre/envs.py

@@ -23,6 +23,7 @@
     VLLM_SPYRE_WORKER_LOG_REDIRECT_DIR: str = ""
     VLLM_SPYRE_GLOO_TIMEOUT_MINUTES: int = 60
     VLLM_SPYRE_REQUIRE_PRECOMPILED_DECODERS: bool = False
+    VLLM_SPYRE_SIMPLE_COMPILE_BACKEND: str = "eager"
 
 logger = init_logger(__name__)
 
@@ -143,7 +144,14 @@ def _backend_backwards_compat() -> str:
     # disable compilation for decoders
     "VLLM_SPYRE_REQUIRE_PRECOMPILED_DECODERS":
     lambda: bool(int(os.getenv("VLLM_SPYRE_REQUIRE_PRECOMPILED_DECODERS", "0"))
-                 )
+                 ),
+
+    # Simple compile backend for some dynamically compiled operations, like
+    # gathering logprobs in the sampler.
+    # Defaults to eager, iductor can be used if python headers and a compiler
+    # are available.
+    "VLLM_SPYRE_SIMPLE_COMPILE_BACKEND":
+    lambda: os.getenv("VLLM_SPYRE_SIMPLE_COMPILE_BACKEND", "eager"),
 }
 # --8<-- [end:env-vars-definition]
 

vllm_spyre/platform.py

@@ -65,9 +65,9 @@ class SpyrePlatform(Platform):
     _num_spyre_blocks_override: int = -1  # override num of KV cache blocks
     _config: VllmConfig = None
 
-    # TODO: see if this needs to be set
+    # Backend for dynamic compilation ops
     # See vllm batched_count_greater_than method
-    # simple_compile_backend: str = "eager"
+    simple_compile_backend: str = envs_spyre.VLLM_SPYRE_SIMPLE_COMPILE_BACKEND
 
     # Needed by vllm/model_executor/layers/pooler.py:562
     current_stream = lambda _: _StreamPlaceholder()