From 9a698f3255a7cc82759f8c623c9ecda50b1e4858 Mon Sep 17 00:00:00 2001 From: Mike G Date: Wed, 22 Jul 2026 22:03:34 -0700 Subject: [PATCH] [Performance][Model] Avoid transient Inkling result allocations (performance, and OOM prevention on smaller memory configurations) (#49487) Signed-off-by: Michael Gschwind Co-authored-by: Michael Gschwind Co-authored-by: OpenAI Codex --- vllm/models/inkling/nvidia/mlp.py | 2 +- vllm/models/inkling/nvidia/moe.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/models/inkling/nvidia/mlp.py b/vllm/models/inkling/nvidia/mlp.py index ab7dec0b70..51deb38e64 100644 --- a/vllm/models/inkling/nvidia/mlp.py +++ b/vllm/models/inkling/nvidia/mlp.py @@ -58,6 +58,6 @@ class InklingDenseMLP(nn.Module): x = silu_and_mul_triton(gate_up) x, _ = self.down_proj(x) if self.global_scale is not None: - x = x * self.global_scale + x.mul_(self.global_scale) # TP-partial output: the layer's reduce-scatter fallback consumes it. return x diff --git a/vllm/models/inkling/nvidia/moe.py b/vllm/models/inkling/nvidia/moe.py index 6d0a550b03..996369f819 100644 --- a/vllm/models/inkling/nvidia/moe.py +++ b/vllm/models/inkling/nvidia/moe.py @@ -528,7 +528,7 @@ class InklingMoE(nn.Module): ) self._routed_sel = None - return out + sink_out + return out.add_(sink_out) # -- weight loading ----------------------------------------------------