2026-05-13 nightly release (03794f5)

Author: pytorchbot

fbgemm_gpu/src/config/feature_gates.cpp

@@ -35,13 +35,13 @@ std::string to_string(const FeatureGateName& value) {
 namespace {
 
 // Returns true iff the env var "FBGEMM_<key>" is set (regardless of value).
-bool env_has_key(const std::string& key) {
+[[maybe_unused]] bool env_has_key(const std::string& key) {
   const auto env_var = "FBGEMM_" + key;
   return std::getenv(env_var.c_str()) != nullptr;
 }
 
 // Reads the env var "FBGEMM_<key>" and returns true iff it is set to "1".
-bool ev_check_key(const std::string& key) {
+bool env_check_key(const std::string& key) {
   const auto env_var = "FBGEMM_" + key;
 
   const auto value = std::getenv(env_var.c_str());
@@ -122,14 +122,14 @@ class FeatureGate {
 #ifdef FBGEMM_FBCODE
     static const NoJkMode mode = NoJkMode::from_env();
     if (mode == NoJkMode::EnvOnly) {
-      value = ev_check_key(key);
+      value = env_check_key(key);
     } else if (mode == NoJkMode::EnvFirstThenJk) {
-      value = env_has_key(key) ? ev_check_key(key) : jk_check_key(key);
+      value = env_has_key(key) ? env_check_key(key) : jk_check_key(key);
     } else /* NoJkMode::JkOnly */ {
       value = jk_check_key(key);
     }
 #else
-    value = ev_check_key(key);
+    value = env_check_key(key);
 #endif
 
     cache_.insert({key, value});

fbgemm_gpu/test/config/feature_gate_no_jk_cpp_test.cpp

@@ -20,7 +20,7 @@ namespace config = fbgemm_gpu::config;
 //
 // In FBCODE builds, these exercise the new EnvFirstThenJk path: env_has_key
 // returning true must short-circuit the JustKnobs lookup, so the value comes
-// from ev_check_key alone.
+// from env_check_key alone.
 //
 // In OSS builds, the same env values flow through the unconditional env-only
 // path (FBGEMM_NO_JK is ignored), and the assertions still hold by

fbgemm_gpu/test/tbe/training/backward_sgd_test.py

@@ -97,8 +97,8 @@ def execute_backward_sgd_(  # noqa C901
             return
         if use_cpu and weights_precision == SparseType.FP16:
             return
-        # No bag ops only work on GPUs, no mixed, no weighted
-        if use_cpu and pooling_mode == PoolingMode.NONE:
+        # V1 API doesn't support nobag on CPU (only PT2 path does)
+        if use_cpu and pooling_mode == PoolingMode.NONE and use_api_v1:
             return
         if mixed and pooling_mode == PoolingMode.NONE:
             return
@@ -449,6 +449,47 @@ def test_backward_sgd(  # noqa C901
             SparseType.FP32,  # output_dtype
         )
 
+    @given(
+        T=st.integers(min_value=1, max_value=3),
+        D=st.sampled_from([2, 4, 128, 256]),
+        B=st.integers(min_value=1, max_value=10),
+        log_E=st.integers(min_value=3, max_value=5),
+        L=st.integers(min_value=1, max_value=20),
+        long_segments=st.booleans(),
+    )
+    @settings(
+        verbosity=VERBOSITY,
+        max_examples=MAX_EXAMPLES,
+        deadline=None,
+        suppress_health_check=[HealthCheck.filter_too_much, HealthCheck.data_too_large],
+    )
+    def test_backward_sgd_fp32_pmNONE_cpu(
+        self,
+        T: int,
+        D: int,
+        B: int,
+        log_E: int,
+        L: int,
+        long_segments: bool,
+    ) -> None:
+        self.execute_backward_sgd_(
+            T,
+            D,
+            B,
+            log_E,
+            L,
+            weights_precision=SparseType.FP32,
+            weighted=False,
+            mixed=False,
+            mixed_B=False,
+            use_cache=False,
+            cache_algorithm=CacheAlgorithm.LRU,
+            long_segments=long_segments,
+            pooling_mode=PoolingMode.NONE,
+            use_cpu=True,
+            output_dtype=SparseType.FP32,
+        )
+
     @given(
         T=st.integers(min_value=1, max_value=5),
         D=st.integers(min_value=2, max_value=256),