diff --git a/containers/vllm/patch-nccl-default-timeout.py b/containers/vllm/patch-nccl-default-timeout.py
new file mode 100644
index 0000000..adc4bed
--- /dev/null
+++ b/containers/vllm/patch-nccl-default-timeout.py
@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+"""Forza il timeout di default di PyTorch per i nuovi process group (usato quando
+timeout=None, es. i gruppi TP/PP/DP/EP creati da vLLM nel worker Ray prima che il
+contesto vllm_config sia disponibile - vedi indagine con debug.py) a un valore piu'
+alto dei 600s hardcoded, invece dei 600s di default per NCCL.
+Uso: patch-nccl-default-timeout.py <path-di-distributed_c10d.py> <secondi>"""
+import sys
+
+path = sys.argv[1]
+seconds = sys.argv[2]
+
+src = open(path).read()
+
+marker = "def _get_default_timeout(backend: Backend) -> timedelta:\n"
+override = (
+    marker
+    + f"    from datetime import timedelta as _td\n"
+    + f"    return _td(seconds={seconds})  # override: vedi patch-nccl-default-timeout.py\n"
+)
+
+count = src.count(marker)
+if count != 1:
+    print(f"ERRORE: marker trovato {count} volte (atteso 1) in {path}", file=sys.stderr)
+    sys.exit(1)
+
+src = src.replace(marker, override, 1)
+open(path, "w").write(src)
+print(f"Patchato {path}: default timeout forzato a {seconds}s")
diff --git a/containers/vllm/vllm-rocm.Containerfile b/containers/vllm/vllm-rocm.Containerfile
index 3b93299..fa5e46c 100644
--- a/containers/vllm/vllm-rocm.Containerfile
+++ b/containers/vllm/vllm-rocm.Containerfile
@@ -100,6 +100,11 @@ RUN RCCL_SRC=""; \
     fi \
     && rm -rf /opt/build-context
 
+# I gruppi paralleli del worker Ray (TP/PP/DP/EP) ricevono timeout=None (vllm_config non ancora disponibile) e finiscono sul default NCCL di PyTorch (600s): lo alziamo qui.
+COPY patch-nccl-default-timeout.py /tmp/patch-nccl-default-timeout.py
+RUN /opt/vllm-rocm/bin/python3 /tmp/patch-nccl-default-timeout.py /opt/vllm-rocm/lib/python3/site-packages/torch/distributed/distributed_c10d.py 3000 \
+    && rm /tmp/patch-nccl-default-timeout.py
+
 # amdsmi non e' importabile di default in questa build, serve nel PYTHONPATH (vedi lemonade-sdk/vllm-rocm#3).
 ENV PYTHONPATH=/opt/vllm-rocm/lib/python3/site-packages/_rocm_sdk_core/share/amd_smi
 
diff --git a/containers/vllm/vllm.container b/containers/vllm/vllm.container
index 693f183..6ec32e1 100644
--- a/containers/vllm/vllm.container
+++ b/containers/vllm/vllm.container
@@ -56,6 +56,8 @@ Environment=TENSOR_PARALLEL_SIZE=2
 Environment=DTYPE=float16
 Environment=KV_CACHE_DTYPE=fp8
 Environment=DISTRIBUTED_TIMEOUT_SECONDS=1800
+# Timeout della singola chiamata worker (es. sample_tokens); default vLLM 300s, letta direttamente da vLLM (non serve modificare l'entrypoint)
+Environment=VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=1800
 # Default vLLM sceglie ROCM_ATTN.
 #Environment=ATTENTION_BACKEND=TRITON_ATTN