Apply gemma's position offset out-of-place instead of in-place (#26799)

python/sglang/srt/environ.py

@@ -199,6 +199,7 @@ class Envs:
     SGLANG_SORT_WEIGHT_FILES = EnvBool(False)
     SGLANG_DISABLED_MODEL_ARCHS = EnvTuple(tuple())
     SGLANG_PREFETCH_BLOCK_SIZE_MB = EnvInt(16)
+    SGLANG_GEMMA_OUT_OF_PLACE_POSITION_MUTATION = EnvBool(False)
 
     # Logging Options
     SGLANG_LOG_GC = EnvBool(False)

python/sglang/srt/models/gemma4_mm.py

@@ -29,6 +29,7 @@ from transformers import (
 )
 
 from sglang.srt.distributed import get_pp_group
+from sglang.srt.environ import envs
 from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
 from sglang.srt.layers.layernorm import Gemma4RMSNorm
 from sglang.srt.layers.linear import ReplicatedLinear
@@ -595,7 +596,11 @@ class Gemma4ForConditionalGeneration(PreTrainedModel):
                 "You must specify exactly one of input_ids or inputs_embeds"
             )
 
-        positions += 1
+        if envs.SGLANG_GEMMA_OUT_OF_PLACE_POSITION_MUTATION.get():
+            positions = positions + 1
+        else:
+            positions += 1
+
         per_layer_inputs = None
         # PLE table and the per-layer projection live on the first rank only,
         # so non-first ranks must skip this and pull per_layer_inputs from the