diff --git a/lib/PTO/Transforms/VMIToVPTO.cpp b/lib/PTO/Transforms/VMIToVPTO.cpp
index 75873dcb7a..eca1494909 100644
--- a/lib/PTO/Transforms/VMIToVPTO.cpp
+++ b/lib/PTO/Transforms/VMIToVPTO.cpp
@@ -5458,6 +5458,51 @@ FailureOr<Value> createIotaContiguousChunk(Location loc, Type resultType,
       .getResult();
 }
 
+// Materialize a contiguous logical iota into a lane-strided physical chunk.
+// On the A5 VPTO path VCI over a narrow floating-point carrier advances in
+// physical lanes.  The lane-stride store then selects every `laneStride`-th
+// lane, so the VCI ramp has to be scaled by the reciprocal stride to preserve
+// a unit logical increment.
+FailureOr<Value> createIotaLaneStrideChunk(
+    Location loc, Type resultType, Value base, int64_t laneStride, int64_t laneOffset, StringAttr orderAttr,
+    PatternRewriter& rewriter)
+{
+    auto vregType = dyn_cast<VRegType>(resultType);
+    if (!vregType || (laneStride != 2 && laneStride != 4)) {
+        return failure();
+    }
+    FailureOr<Value> mask = createAllTrueMaskForVReg(loc, vregType, rewriter);
+    FailureOr<Value> zero = createScalarOffsetConstant(loc, base.getType(), 0, rewriter);
+    // This materialization is needed for narrow floating-point carriers (the
+    // f16 PK_B32 case in particular).  Integer VCI uses its native unit-lane
+    // semantics and must not be scaled here.
+    int64_t scale = 1;
+    if (isa<FloatType>(base.getType()))
+        scale = laneStride;
+    FailureOr<Value> factor;
+    if (isa<FloatType>(base.getType())) {
+        auto floatType = cast<FloatType>(base.getType());
+        factor = rewriter.create<arith::ConstantOp>(loc, rewriter.getFloatAttr(floatType, 1.0 / scale)).getResult();
+    } else {
+        factor = createScalarOffsetConstant(loc, base.getType(), 1, rewriter);
+    }
+    if (failed(mask) || failed(zero) || failed(factor)) {
+        return failure();
+    }
+    Value indices = rewriter.create<VciOp>(loc, resultType, *zero, StringAttr{}).getResult();
+    Value scaled = rewriter.create<VmulsOp>(loc, resultType, indices, *factor, *mask).getResult();
+    StringRef order = orderAttr ? orderAttr.getValue() : StringRef("ASC");
+    FailureOr<Value> chunkBase = createIotaChunkBase(loc, base, laneOffset, order, rewriter);
+    if (failed(chunkBase)) {
+        return failure();
+    }
+    if (order == "DESC") {
+        Value baseVector = rewriter.create<VdupOp>(loc, resultType, *chunkBase, *mask, nullptr).getResult();
+        return rewriter.create<VsubOp>(loc, resultType, baseVector, scaled, *mask).getResult();
+    }
+    return rewriter.create<VaddsOp>(loc, resultType, scaled, *chunkBase, *mask).getResult();
+}
+
 /// Pack group-periodic ramps inside one physical VL when S < physVL and
 /// physVL % S == 0 (e.g. i32 L=64,group=2 → [base..base+31 | base..base+31]).
 ///
@@ -5751,6 +5796,28 @@ struct OneToNVMIIotaOpPattern : OneToNOpConversionPattern<IotaOp> {
       return success();
     }
 
+    if (layout.isContiguous() && layout.getLaneStride() != 1) {
+        int64_t laneStride = layout.getLaneStride();
+        if (laneStride != 2 && laneStride != 4) {
+            return rewriter.notifyMatchFailure(op, "unsupported contiguous iota lane_stride");
+        }
+        int64_t logicalLanesPerChunk = *lanesPerPart / laneStride;
+        for (auto [index, resultType] : llvm::enumerate(resultTypes)) {
+            if (!isa<VRegType>(resultType)) {
+                return rewriter.notifyMatchFailure(op, "iota result must be vreg");
+            }
+            FailureOr<Value> result = createIotaLaneStrideChunk(
+                op.getLoc(), resultType, *base, laneStride, static_cast<int64_t>(index) * logicalLanesPerChunk,
+                op.getOrderAttr(), rewriter);
+            if (failed(result)) {
+                return rewriter.notifyMatchFailure(op, "failed to materialize lane-strided iota chunk");
+            }
+            results.push_back(*result);
+        }
+        replaceOpWithFlatConvertedValues(rewriter, op, results, *this->getTypeConverter());
+        return success();
+    }
+
     if (layout.isContiguous()) {
       for (auto [index, resultType] : llvm::enumerate(resultTypes)) {
         if (!isa<VRegType>(resultType))
diff --git a/packaging/ptoas-vmi/pyproject.toml.patch b/packaging/ptoas-vmi/pyproject.toml.patch
index f2f71ec41c..87dc5d8c72 100644
--- a/packaging/ptoas-vmi/pyproject.toml.patch
+++ b/packaging/ptoas-vmi/pyproject.toml.patch
@@ -10,7 +10,7 @@ diff --git a/pyproject.toml b/pyproject.toml
 +version = "0.1.6"
 +description = "PTO Assembler & Optimizer with VMI support"
  readme = "README.md"
- requires-python = ">=3.10"
+ requires-python = ">=3.7"
  license = "Apache-2.0"
 @@ -43,4 +43,5 @@ cmake.version = "CMakeLists.txt"
  ninja.version = ">=1.10"
diff --git a/test/lit/vmi_new/vmi_to_vpto_iota.pto b/test/lit/vmi_new/vmi_to_vpto_iota.pto
index 79f8443c48..8652c76170 100644
--- a/test/lit/vmi_new/vmi_to_vpto_iota.pto
+++ b/test/lit/vmi_new/vmi_to_vpto_iota.pto
@@ -49,6 +49,19 @@ module {
     return %part : !pto.vreg<128xi16>
   }
 
+  // Regression for #1411: a compact f16 store chooses a lane_stride=2
+  // carrier.  The physical ramp must advance by two so PK_B32 observes a
+  // logical step of one.
+  func.func @vmi_to_vpto_iota_f16_lane_stride2(%base: f16)
+      -> !pto.vreg<128xf16> {
+    %value = pto.vmi.vci %base
+        : f16 -> !pto.vmi.vreg<64xf16, #pto.vmi.layout<contiguous, lane_stride = 2>>
+    %part = "pto.vmi.unpack"(%value)
+        : (!pto.vmi.vreg<64xf16, #pto.vmi.layout<contiguous, lane_stride = 2>>)
+        -> !pto.vreg<128xf16>
+    return %part : !pto.vreg<128xf16>
+  }
+
   func.func @vmi_to_vpto_iota_f16_deint2_asc(%base: f16)
       -> (!pto.vreg<128xf16>, !pto.vreg<128xf16>) {
     %value = pto.vmi.vci %base
@@ -101,6 +114,15 @@ module {
 // CHECK-NOT: !pto.vmi.
 // CHECK-NOT: unrealized_conversion_cast
 
+// CHECK-LABEL: func.func @vmi_to_vpto_iota_f16_lane_stride2(
+// CHECK: %[[C2:.*]] = arith.constant 5.000000e-01 : f16
+// CHECK: %[[C0:.*]] = arith.constant 0.000000e+00 : f16
+// CHECK: %[[IDX:.*]] = pto.vci %[[C0]] : f16 -> !pto.vreg<128xf16>
+// CHECK: %[[SCALED:.*]] = pto.vmuls %[[IDX]], %[[C2]]
+// CHECK: %[[RESULT:.*]] = pto.vadds %[[SCALED]], %arg0
+// CHECK: return %[[RESULT]]
+// CHECK-NOT: pto.vmi.
+
 // CHECK-LABEL: func.func @vmi_to_vpto_iota_f16_deint2_asc(
 // CHECK: arith.constant 1.000000e+00 : f16
 // CHECK: arith.constant 2.000000e+00 : f16