adaption DYNAMO for DCU

components/planner/src/dynamo/planner/local_connector.py

@@ -148,7 +148,7 @@ class LocalConnector(PlannerConnector):
             if not available_gpus:
                 raise ValueError("No GPUs available for allocation")
             gpu_id = available_gpus[0]
-            watcher_env["CUDA_VISIBLE_DEVICES"] = gpu_id
+            watcher_env["HIP_VISIBLE_DEVICES"] = gpu_id
 
         watcher_env["DYNAMO_SERVICE_CONFIG"] = service_config
 

deploy/sdk/src/dynamo/sdk/cli/allocator.py

@@ -207,7 +207,7 @@ class ResourceAllocator:
 
                 # Generate environment variables for each worker
                 for _ in range(num_workers):
-                    env_vars = {"CUDA_VISIBLE_DEVICES": ",".join(map(str, assigned))}
+                    env_vars = {"HIP_VISIBLE_DEVICES": ",".join(map(str, assigned))}
                     resource_envs.append(env_vars)
             else:
                 logger.info(
@@ -221,7 +221,7 @@ class ResourceAllocator:
                     )
 
                     # Generate environment variables for this worker
-                    env_vars = {"CUDA_VISIBLE_DEVICES": ",".join(map(str, assigned))}
+                    env_vars = {"HIP_VISIBLE_DEVICES": ",".join(map(str, assigned))}
 
                     # If we have comprehensive GPU stats, log them
                     try:
@@ -242,7 +242,8 @@ class ResourceAllocator:
                         logger.debug(f"Failed to get GPU stats: {e}")
 
                     resource_envs.append(env_vars)
-
+        # else:
+        #     resource_envs = config["envs"]
         logger.info(
             f"Final resource allocation - workers: {num_workers}, envs: {resource_envs}"
         )

examples/llm/configs/agg.yaml

@@ -13,12 +13,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 Common:
-  model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  block-size: 64
+  model: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
+  block-size: 16
   max-model-len: 16384
 
 Frontend:
-  served_model_name: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+  served_model_name: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
   endpoint: dynamo.Processor.chat/completions
   port: 8000
 

examples/llm/configs/disagg.yaml

@@ -13,13 +13,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 Common:
-  model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  block-size: 64
+  model: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
+  block-size: 16
   max-model-len: 16384
   kv-transfer-config: '{"kv_connector":"DynamoNixlConnector"}'
 
 Frontend:
-  served_model_name: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+  served_model_name: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
   endpoint: dynamo.Processor.chat/completions
   port: 8000
 
@@ -29,21 +29,26 @@ Processor:
 
 VllmWorker:
   remote-prefill: true
-  conditional-disagg: true
+  conditional-disagg: false
   max-local-prefill-length: 10
-  max-prefill-queue-size: 2
+  max-prefill-queue-size: 64
+  tensor-parallel-size: 1
+  enable-prefix-caching: false
   ServiceArgs:
     workers: 1
     resources:
       gpu: 1
+
   common-configs: [model, block-size, max-model-len, kv-transfer-config]
 
 PrefillWorker:
   max-num-batched-tokens: 16384
+  tensor-parallel-size: 1
   ServiceArgs:
     workers: 1
     resources:
       gpu: 1
+
   common-configs: [model, block-size, max-model-len, kv-transfer-config]
 
 Planner:

examples/llm/configs/disagg_router.yaml

@@ -13,14 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 Common:
-  model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-  block-size: 64
+  model: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
+  block-size: 16
   max-model-len: 16384
   router: kv
   kv-transfer-config: '{"kv_connector":"DynamoNixlConnector"}'
 
 Frontend:
-  served_model_name: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+  served_model_name: /models/deepseek-r1/DeepSeek-R1-Distill-Llama-8B
   endpoint: dynamo.Processor.chat/completions
   port: 8000
 
@@ -34,9 +34,9 @@ Router:
 VllmWorker:
   max-num-batched-tokens: 16384
   remote-prefill: true
-  conditional-disagg: true
+  conditional-disagg: false
   max-local-prefill-length: 10
-  max-prefill-queue-size: 2
+  max-prefill-queue-size: 64
   tensor-parallel-size: 1
   enable-prefix-caching: true
   ServiceArgs:

examples/llm/configs/mutinode_disagg_r1.yaml

@@ -14,7 +14,7 @@
 # limitations under the License.
 Common:
   model: deepseek-ai/DeepSeek-R1
-  block-size: 64
+  block-size: 16
   max-model-len: 16384
   kv-transfer-config: '{"kv_connector":"DynamoNixlConnector"}'
   tensor-parallel-size: 16

launch/llmctl/src/main.rs

@@ -228,9 +228,9 @@ async fn add_model(
         endpoint_name
     );
 
-    if model_name.starts_with('/') {
-        raise!("Model name '{}' cannot start with a slash", model_name);
-    }
+    // if model_name.starts_with('/') {
+    //     raise!("Model name '{}' cannot start with a slash", model_name);
+    // }
 
     let parts: Vec<&str> = endpoint_name.split('.').collect();