feat: Enable intel gaudi on dynamo (#4209)

Signed-off-by: Spycsh <sihan.chen@intel.com>

components/src/dynamo/vllm/handlers.py

@@ -389,6 +389,16 @@ class PrefillWorkerHandler(BaseWorkerHandler):
         sampling_params.extra_args["kv_transfer_params"] = {
             "do_remote_decode": True,
         }
+        sampling_params_defaults = {
+            "do_remote_prefill": False,
+            "remote_engine_id": None,
+            "remote_block_ids": None,
+            "remote_host": None,
+            "remote_port": None,
+        }
+        # Add only missing keys
+        for k, v in sampling_params_defaults.items():
+            sampling_params.extra_args["kv_transfer_params"].setdefault(k, v)
         # Override for prefill: only generate 1 token
         sampling_params.max_tokens = 1
         sampling_params.min_tokens = 1

examples/backends/vllm/launch/disagg_router_gaudi.sh

+#!/bin/bash
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+set -e
+trap 'echo Cleaning up...; kill 0' EXIT
+
+# Set deterministic hash for KV event IDs
+export PYTHONHASHSEED=0
+
+# Common configuration
+MODEL="Qwen/Qwen3-0.6B"
+BLOCK_SIZE=64
+VLLM_NIXL_DEVICE_TO_DEVICE=false
+VLLM_SKIP_WARMUP=true
+PT_HPU_LAZY_MODE=0
+NIXL_BUFFER_DEVICE=cpu
+VLLM_NIXL_BACKEND=UCX
+
+
+# Start frontend with KV routing
+# The frontend will automatically detect prefill workers and activate an internal prefill router
+# edit --router-mode to random / round-robin / kv
+python -m dynamo.frontend \
+    --router-mode kv \
+    --http-port 8000 \
+    --router-reset-states &
+
+# two decode workers
+# --enforce-eager is added for quick deployment. for production use, need to remove this flag
+VLLM_NIXL_SIDE_CHANNEL_PORT=20096 \
+HABANA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \
+    --model $MODEL \
+    --block-size $BLOCK_SIZE \
+    --kv-transfer-config "{\"kv_connector\": \"NixlConnector\", \"kv_role\": \"kv_both\", \"kv_buffer_device\": \"${NIXL_BUFFER_DEVICE}\", \"kv_connector_extra_config\": {\"backends\": [\"${VLLM_NIXL_BACKEND}\"]}}" \
+    --connector none \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5556", "enable_kv_cache_events":true}' &
+
+VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \
+HABANA_VISIBLE_DEVICES=1 python3 -m dynamo.vllm \
+    --model $MODEL \
+    --block-size $BLOCK_SIZE \
+    --kv-transfer-config "{\"kv_connector\": \"NixlConnector\", \"kv_role\": \"kv_both\", \"kv_buffer_device\": \"${NIXL_BUFFER_DEVICE}\", \"kv_connector_extra_config\": {\"backends\": [\"${VLLM_NIXL_BACKEND}\"]}}" \
+    --connector none \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557", "enable_kv_cache_events":true}' &
+
+# two prefill workers
+# When registered with --is-prefill-worker, these workers are automatically detected
+# by the frontend, which activates an internal prefill router for KV-aware prefill routing
+VLLM_NIXL_SIDE_CHANNEL_PORT=20098 \
+HABANA_VISIBLE_DEVICES=2 python3 -m dynamo.vllm \
+    --model $MODEL \
+    --block-size $BLOCK_SIZE \
+    --kv-transfer-config "{\"kv_connector\": \"NixlConnector\", \"kv_role\": \"kv_both\", \"kv_buffer_device\": \"${NIXL_BUFFER_DEVICE}\", \"kv_connector_extra_config\": {\"backends\": [\"${VLLM_NIXL_BACKEND}\"]}}" \
+    --connector none \
+    --is-prefill-worker \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5558", "enable_kv_cache_events":true}'&
+
+VLLM_NIXL_SIDE_CHANNEL_PORT=20099 \
+HABANA_VISIBLE_DEVICES=3 python3 -m dynamo.vllm \
+    --model $MODEL \
+    --block-size $BLOCK_SIZE \
+    --kv-transfer-config "{\"kv_connector\": \"NixlConnector\", \"kv_role\": \"kv_both\", \"kv_buffer_device\": \"${NIXL_BUFFER_DEVICE}\", \"kv_connector_extra_config\": {\"backends\": [\"${VLLM_NIXL_BACKEND}\"]}}" \
+    --connector none \
+    --is-prefill-worker \
+    --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5559", "enable_kv_cache_events":true}'