From 0af29d957a57b436c5348363befcbb58755f05d8 Mon Sep 17 00:00:00 2001
From: turboderp <11859846+turboderp@users.noreply.github.com>
Date: Wed, 15 Oct 2025 10:40:19 +0200
Subject: [PATCH] Fix #390

---
 backends/exllamav2/model.py | 11 +++--------
 backends/exllamav3/model.py | 13 +++++--------
 2 files changed, 8 insertions(+), 16 deletions(-)

diff --git a/backends/exllamav2/model.py b/backends/exllamav2/model.py
index 4394697..dead5f5 100644
--- a/backends/exllamav2/model.py
+++ b/backends/exllamav2/model.py
@@ -1336,14 +1336,9 @@ class ExllamaV2Container(BaseModelContainer):
         negative_context_len = input_ids[1].size(dim=-1) if negative_prompt else 0
 
         # Automatically set max_tokens to fill up the context
-        # This should be an OK default, but may be changed in the future
-        max_tokens = unwrap(
-            params.max_tokens,
-            self.config.max_seq_len - max(context_len, negative_context_len),
-        )
-        if max_tokens < 1:
-            logger.warning("max_tokens must be a positive integer, setting to 1.")
-            max_tokens = 1
+        max_tokens = unwrap(params.max_tokens, 0)
+        if max_tokens <= 0:
+            max_tokens = self.config.max_seq_len - max(context_len, negative_context_len)
 
         # Determine if the negative context or the context length is bigger
         context_to_check = max(negative_context_len, context_len)
diff --git a/backends/exllamav3/model.py b/backends/exllamav3/model.py
index c1277d0..d385de7 100644
--- a/backends/exllamav3/model.py
+++ b/backends/exllamav3/model.py
@@ -951,14 +951,11 @@ class ExllamaV3Container(BaseModelContainer):
         # The first index will always be the positive prompt
         context_len = input_ids[0].size(dim=-1)
 
-        # Automatically set max_tokens to fill up the context
-        max_tokens = unwrap(
-            params.max_tokens if params.max_tokens > 0 else None,
-            self.max_seq_len - context_len - 1,
-        )
-        if max_tokens < 1:
-            logger.warning("max_tokens must be a positive integer, setting to 1.")
-            max_tokens = 1
+        # Unless specified in the request, automatically set max_tokens to fill up
+        # the context
+        max_tokens = unwrap(params.max_tokens, 0)
+        if max_tokens <= 0:
+            max_tokens = self.max_seq_len - context_len - 1
 
         # Check total length of prompt against max context length
         if context_len > self.max_seq_len: