From 6b2bf66cd9cd0448b0e9f3af8a54e9e10686fdf2 Mon Sep 17 00:00:00 2001
From: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
Date: Mon, 13 Apr 2026 15:10:07 -0500
Subject: [PATCH] fix[glm4.7 flash]: properly detect `gfx95_quant_format`
 (#22720)

---
 python/sglang/srt/models/glm4_moe_lite.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/python/sglang/srt/models/glm4_moe_lite.py b/python/sglang/srt/models/glm4_moe_lite.py
index 03ff3d201..5b08f1f45 100644
--- a/python/sglang/srt/models/glm4_moe_lite.py
+++ b/python/sglang/srt/models/glm4_moe_lite.py
@@ -403,6 +403,8 @@ class Glm4MoeLiteDecoderLayer(DeepseekV2DecoderLayer):
             config.hidden_size, eps=config.rms_norm_eps
         )
 
+        self._gfx95_quant_format = self._detect_gfx95_quant_format()
+
         self.layer_communicator = LayerCommunicator(
             layer_scatter_modes=self.layer_scatter_modes,
             input_layernorm=self.input_layernorm,