diff --git a/crates/luminal_cuda_lite/src/ops/cublaslt/device_call.rs b/crates/luminal_cuda_lite/src/ops/cublaslt/device_call.rs
index facdb25..7a4629a 100644
--- a/crates/luminal_cuda_lite/src/ops/cublaslt/device_call.rs
+++ b/crates/luminal_cuda_lite/src/ops/cublaslt/device_call.rs
@@ -116,6 +116,7 @@ impl Layout {
     fn new(desc: &LtDesc, dtype: luminal::dtype::PlanDtype) -> Result<Self> {
         let dtype = match dtype {
             luminal::dtype::PlanDtype::F32 => sys::cudaDataType_t::CUDA_R_32F,
+            luminal::dtype::PlanDtype::F16 => sys::cudaDataType_t::CUDA_R_16F,
             luminal::dtype::PlanDtype::Bf16 => sys::cudaDataType_t::CUDA_R_16BF,
             dtype => return Err(anyhow!("cuBLASLt unsupported matrix dtype {dtype:?}")),
         };
diff --git a/crates/luminal_cuda_lite/src/ops/cublaslt/egg/cublaslt_marker_site.egg b/crates/luminal_cuda_lite/src/ops/cublaslt/egg/cublaslt_marker_site.egg
index dd8287f..de7cf58 100644
--- a/crates/luminal_cuda_lite/src/ops/cublaslt/egg/cublaslt_marker_site.egg
+++ b/crates/luminal_cuda_lite/src/ops/cublaslt/egg/cublaslt_marker_site.egg
@@ -69,3 +69,33 @@
   :name "cuda_lite/cublaslt/egg/cublaslt_marker_site: BF16 matmul with F32 output"
 )
 ; ============================================================================
+
+; FP16 storage with explicit FP32 arithmetic. Widening FP16 to FP32 is
+; exact; an FP16 product fits in FP32. Consume the original storage while
+; retaining the F32 result and reduction contract. This is NOT a rule for
+; cast(F16 matmul, F32): that graph rounds products/reduction in F16.
+(rule
+  (
+    (= ?out (LogicalHelperMatrixMultiply ?aw ?bw))
+    (= ?aw (LogicalCast ?a (F32)))
+    (= ?bw (LogicalCast ?b (F32)))
+    (= (dtype-of ?a) (F16))
+    (= (dtype-of ?b) (F16))
+    (= (dtype-of ?out) (F32))
+  )
+  ((let ?site (CublasLtLogicalMatmulSite ?a ?b ?out)))
+  :ruleset backend
+  :name "cuda_lite/cublaslt/egg/cublaslt_marker_site: FP16 storage with explicit FP32 arithmetic"
+)
+
+; Let the existing F32 transpose-sandwich rule expose the same exact
+; widening on a transposed operand. Transposition changes indices only.
+(rule
+  (
+    (= ?out (LogicalHelperMatrixTranspose (LogicalCast ?a (F32))))
+    (= (dtype-of ?a) (F16))
+  )
+  ((union ?out (LogicalCast (LogicalHelperMatrixTranspose ?a) (F32))))
+  :ruleset backend
+  :name "cuda_lite/cublaslt/egg/cublaslt_marker_site: transpose commutes with FP16 widening"
+)