From 092387963c09ff3dde648e4ff620174d5d32c5d3 Mon Sep 17 00:00:00 2001
From: Jimmy Lee <58957694+thisisjimmyfb@users.noreply.github.com>
Date: Fri, 10 Jul 2026 19:05:35 -0700
Subject: [PATCH] [BugFix] weights processing peak memory reduction for nvfp4
 MoE layers (#46276)

Signed-off-by: Jimmy Lee <hirejimmylee@gmail.com>
---
 tests/kernels/moe/test_flashinfer_b12x_moe.py | 33 +++++++------------
 .../quantization/utils/flashinfer_fp4_moe.py  | 32 +++++++++++++-----
 2 files changed, 36 insertions(+), 29 deletions(-)

diff --git a/tests/kernels/moe/test_flashinfer_b12x_moe.py b/tests/kernels/moe/test_flashinfer_b12x_moe.py
index b15cbcdd812..d1859e04f68 100644
--- a/tests/kernels/moe/test_flashinfer_b12x_moe.py
+++ b/tests/kernels/moe/test_flashinfer_b12x_moe.py
@@ -41,6 +41,9 @@ from vllm.model_executor.layers.fused_moe.config import nvfp4_moe_quant_config
 from vllm.model_executor.layers.fused_moe.experts.flashinfer_b12x_moe import (
     FlashInferB12xExperts,
 )
+from vllm.model_executor.layers.quantization.utils.flashinfer_fp4_moe import (
+    reorder_w1w3_to_w3w1,
+)
 from vllm.utils.torch_utils import set_random_seed
 
 # Dimensions chosen to satisfy FP4 alignment requirements (k multiple of 256,
@@ -53,23 +56,6 @@ MNK_FACTORS = [
 ]
 
 
-def _reorder_gate_up_to_up_gate(
-    w: torch.Tensor,
-    w_s: torch.Tensor,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Swap gate and up-projection halves along dim=1 to [up, gate] order.
-
-    The B12x kernel expects weights in [up (w3), gate (w1)] order while the
-    BF16 reference uses [gate (w1), up (w3)].  This replicates the reordering
-    done at model-load time by ``prepare_nvfp4_moe_layer_for_fi_or_cutlass``.
-    """
-    n = w.shape[1] // 2
-    return (
-        torch.cat([w[:, n:, :], w[:, :n, :]], dim=1),
-        torch.cat([w_s[:, n:, :], w_s[:, :n, :]], dim=1),
-    )
-
-
 def _process_b12x_weights(
     experts: FlashInferB12xExperts,
     w1_scale: torch.Tensor,
@@ -142,9 +128,14 @@ def test_flashinfer_b12x_moe(
         sf_vec_size = 16
 
         # W1: reorder BF16 from [gate, up] → [up, gate], then quantise.
-        w1_reordered = torch.cat(
-            [w1_bf16[:, n:, :], w1_bf16[:, :n, :]], dim=1
-        )  # shape (e, 2n, k), [up, gate]
+        # Note: in reorder_w1w3_to_w3w1, "w1" refers to the gate projection
+        # and "w3" refers to the up projection.
+        # A dummy scale is passed and discarded; real scales come from
+        # fp4_quantize after reordering.
+        w1_reordered, _ = reorder_w1w3_to_w3w1(
+            w1_bf16.clone(),
+            torch.ones((e, 2 * n, 1), device="cuda", dtype=torch.float32),
+        )
         w1_flat = w1_reordered.reshape(e * 2 * n, k)
         w1_q_flat, w1_sf_flat = fp4_quantize(
             w1_flat,
@@ -190,6 +181,7 @@ def test_flashinfer_b12x_moe(
             moe_config=moe_config,
             quant_config=quant_config,
         )
+
         _process_b12x_weights(
             experts,
             w1_blockscale,
@@ -324,7 +316,6 @@ def test_flashinfer_b12x_moe_relu2(
                 use_monolithic=False,
             ),
             experts,
-            inplace=False,
         )
 
         score = torch.randn((m, e), device="cuda", dtype=dtype)
diff --git a/vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py b/vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
index 23a7131a582..2fd21c2dff8 100644
--- a/vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
+++ b/vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
@@ -23,7 +23,6 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
-
 __all__ = [
     "reorder_w1w3_to_w3w1",
 ]
@@ -32,18 +31,35 @@ __all__ = [
 def reorder_w1w3_to_w3w1(
     weight: torch.Tensor, scale: torch.Tensor, dim: int = -2
 ) -> tuple[torch.Tensor, torch.Tensor]:
-    """Re-order the concatenated `[w1, w3]` tensors to `[w3, w1]`"""
+    """Re-order concatenated `[w1, w3]` tensors to `[w3, w1]` in-place.
+
+    `weight` and `scale` must be contiguous; they remain contiguous on return.
+    """
+    assert weight.is_contiguous(), "weight must be contiguous"
+    assert scale.is_contiguous(), "scale must be contiguous"
     size = weight.size(dim)
     assert size % 2 == 0, f"Expected even size in dim {dim}, got {size}"
     half = size // 2
+    d = dim % weight.dim()
 
-    w1, w3 = weight.split(half, dim=dim)
-    s1, s3 = scale.split(half, dim=dim)
-
-    return (
-        torch.cat([w3, w1], dim=dim).contiguous(),
-        torch.cat([s3, s1], dim=dim).contiguous(),
+    # 64 MB transient cap
+    bytes_per_row = max(
+        weight.numel() // size * weight.element_size(),
+        scale.numel() // size * scale.element_size(),
     )
+    chunk = max(1, min(half, (64 << 20) // max(bytes_per_row, 1)))
+
+    fa, fb = [slice(None)] * weight.dim(), [slice(None)] * weight.dim()
+    for off in range(0, half, chunk):
+        end = min(off + chunk, half)
+        fa[d], fb[d] = slice(off, end), slice(half + off, half + end)
+        a, b = tuple(fa), tuple(fb)
+        for t in (weight, scale):
+            tmp = t[b].clone()
+            t[b] = t[a]
+            t[a] = tmp
+
+    return weight, scale
 
 
 def interleave_linear_and_gate(