From a3440098a4d4fd0daf4f2c2e5091647c6f60014b Mon Sep 17 00:00:00 2001
From: turboderp <11859846+turboderp@users.noreply.github.com>
Date: Tue, 29 Apr 2025 20:44:10 +0200
Subject: [PATCH] Add Qwen3ForCausalLM

---
 exllamav2/architecture.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/exllamav2/architecture.py b/exllamav2/architecture.py
index 300fe0c..1797dc4 100644
--- a/exllamav2/architecture.py
+++ b/exllamav2/architecture.py
@@ -428,6 +428,19 @@ class ExLlamaV2ArchParams:
             self.lm.attention_bias_qkv = True
             self.lm.supports_tp = True
 
+        # Qwen3
+
+        if arch_string == "Qwen3ForCausalLM":
+            arch_recognized = True
+            self.lm.layer_keys += \
+                layer_keys_llama_norms + \
+                layer_keys_llama_attn + \
+                layer_keys_llama_mlp
+            self.lm.expect_keys += \
+                expect_keys_llama
+            self.lm.supports_tp = True
+            self.lm.default_use_qk_norm = True
+
         # Qwen2-VL (2, 2.5)
 
         if arch_string in ["Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration"]: