feat: Add support for disaggregation with pp with pytorch backend (#6369)

Signed-off-by: Patrice Castonguay <[email protected]>
Signed-off-by: raayandhar <[email protected]>
Signed-off-by: Lizhi Zhou <[email protected]>
Signed-off-by: pcastonguay <[email protected]>
Co-authored-by: raayandhar <[email protected]>
Co-authored-by: Lizhi Zhou <[email protected]>
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

Author: 4 people

cpp/tensorrt_llm/batch_manager/cacheFormatter.cpp

@@ -840,6 +840,8 @@ void CacheFormatter::unformat(TransferSession& session)
     if (selfConfig.getModelConfig().mNbKvHeadsPerLayer.size() != destConfig.getModelConfig().mNbKvHeadsPerLayer.size())
     {
         TLLM_LOG_WARNING("CacheFormatter::inquireSupport: only support same number of layers");
+        TLLM_LOG_WARNING("self: %zu dest %zu", selfConfig.getModelConfig().mNbKvHeadsPerLayer.size(),
+            destConfig.getModelConfig().mNbKvHeadsPerLayer.size());
         return false;
     }
     int selfNumLayers = selfConfig.getModelConfig().mNbKvHeadsPerLayer.size();

scripts/build_wheel.py

@@ -71,7 +71,10 @@ def clear_folder(folder_path):
         if os.path.isdir(item_path) and not os.path.islink(item_path):
             rmtree(item_path)
         else:
-            os.remove(item_path)
+            try:
+                os.remove(item_path)
+            except (OSError, IOError) as e:
+                print(f"Failed to remove {item_path}: {e}", file=sys.stderr)
 
 
 def sysconfig_scheme(override_vars=None):

tensorrt_llm/_torch/pyexecutor/kv_cache_transceiver.py

@@ -96,13 +96,13 @@ def __init__(self, mapping: Mapping, kv_cache_manager: KVCacheManager,
                  attention_type: AttentionTypeCpp,
                  cache_transceiver_config: CacheTransceiverConfig):
         world_config = mapping_to_world_config(mapping)
-        num_kv_heads_per_layer = kv_cache_manager.num_kv_heads_per_layer
+        total_num_kv_heads_per_layer = kv_cache_manager.total_num_kv_heads_per_layer
         head_dim = kv_cache_manager.head_dim
         tokens_per_block = kv_cache_manager.tokens_per_block
         dtype = kv_cache_manager.dtype
 
         self.impl = CacheTransceiverCpp(kv_cache_manager.impl,
-                                        num_kv_heads_per_layer, head_dim,
+                                        total_num_kv_heads_per_layer, head_dim,
                                         tokens_per_block, world_config, dtype,
                                         attention_type,
                                         cache_transceiver_config)

tensorrt_llm/_torch/pyexecutor/py_executor.py

@@ -122,6 +122,7 @@ class BatchState:
 @dataclasses.dataclass
 class BatchStatePP(BatchState):
     microbatch_id: int = -1
+    scheduled_ctx_reqs: list[LlmRequest] = None
 
 
 class PyExecutor:
@@ -641,6 +642,7 @@ def _need_return_log_probs(self, scheduled_requests: ScheduledRequests):
         return False
 
     def _executor_loop_pp(self):
+        logger.debug(f"Starting executor loop for pp_rank {self.dist.pp_rank}")
         torch.cuda.set_device(self.device_id)
         microbatch_id = 0
         with self._profiler() as profile_step:
@@ -654,6 +656,9 @@ def _executor_loop_pp(self):
                 if self.should_stop_processing:
                     break
 
+                if self.kv_cache_transceiver:
+                    self._check_disagg_gen_transfer_status()
+
                 if self.enable_iter_perf_stats:
                     iter_stats = self._get_init_iter_stats(
                         len(new_requests),
@@ -662,9 +667,23 @@ def _executor_loop_pp(self):
 
                 self._pad_attention_dp_dummy_request()
 
-                scheduled_batch, _, _ = self._schedule()
+                scheduled_batch, fitting_disagg_gen_init_requests, num_fitting_reqs = self._schedule(
+                )
+
+                if self.kv_cache_transceiver:
+                    # For requests that are fitting disagg gen init, also prepare resources for KV cache manager
+                    self._prepare_disagg_gen_init(
+                        fitting_disagg_gen_init_requests)
+
+                    if num_fitting_reqs == 0 and not fitting_disagg_gen_init_requests:
+                        logger.warning(
+                            "num_fitting_reqs=0 and fitting_disagg_gen_init_requests is empty, may not have enough kvCache"
+                        )
+                        self.kv_cache_transceiver.check_context_transfer_status(
+                            1)
 
                 self.num_scheduled_requests = scheduled_batch.batch_size
+
                 logger.debug(
                     f'has {len(self.active_requests)} active_request, '
                     f'scheduled {len(scheduled_batch.context_requests)} context requests and '
@@ -677,7 +696,7 @@ def _executor_loop_pp(self):
                     can_queue = 0 not in tp_batch_sizes
                 else:
                     can_queue = scheduled_batch.batch_size > 0
-                    if not can_queue:
+                    if not can_queue and not self.kv_cache_transceiver:
                         assert len(self.inflight_req_ids) > 0, (
                             "fail to schedule any pending request, probably run out of resource"
                         )
@@ -686,8 +705,28 @@ def _executor_loop_pp(self):
                     self.micro_batches[microbatch_id] = None
                 else:
                     self._add_inflight_ids(scheduled_batch)
+
+                    if self.kv_cache_transceiver:
+                        # For generation requests which have completed KV cache transfer
+                        self._prepare_disagg_gen_transmission_complete(
+                            scheduled_batch)
+
                     self.resource_manager.prepare_resources(scheduled_batch)
 
+                    # The generation requests that are do not have batch_idx,
+                    # needs to be in front of the batch due to the assumptions
+                    # made in model_engine.py::_forward_step. This is only important
+                    # for disaggregated serving. For non-disaggregated serving,
+                    # the generation requests always have batch_idx.
+                    scheduled_batch.generation_requests = sorted(  # stable sort
+                        scheduled_batch.generation_requests,
+                        key=lambda req: int(req.py_batch_idx is not None),
+                    )
+
+                    if self.kv_cache_transceiver:
+                        # Return the first token to the client
+                        self._handle_first_token_response(scheduled_batch)
+
                     # Stage 1: Async forward (all ranks) and decoding pass (last rank only)
                     if not self.dist.is_last_pp_rank:
                         sample_state = self._forward_step_inter_pp(
@@ -715,6 +754,7 @@ def _executor_loop_pp(self):
                         iter_start_time=iter_start_time,
                         iter_stats=iter_stats,
                         microbatch_id=microbatch_id,
+                        scheduled_ctx_reqs=scheduled_batch.context_requests,
                     )
 
                     self.micro_batches[microbatch_id] = batch_state
@@ -779,6 +819,11 @@ def _executor_loop_pp(self):
                 if previous_batch is not None:
                     with torch.cuda.nvtx.range("_handle_previous_batch_pp"):
                         self._update_requests(previous_batch.sample_state)
+
+                        if self.kv_cache_transceiver and previous_batch.scheduled_ctx_reqs:
+                            self._send_disagg_ctx_cache(
+                                previous_batch.scheduled_ctx_reqs)
+
                         self._handle_canceled_requests()
                         finished_requests = self._handle_responses()
                         previous_scheduled_batch = previous_batch.sample_state.scheduled_requests
@@ -787,6 +832,9 @@ def _executor_loop_pp(self):
                         self._remove_inflight_ids(previous_scheduled_batch)
                     self.micro_batches[prev_microbatch_id] = None
 
+                if self.kv_cache_transceiver and self.ctx_in_transmission_requests:
+                    self._terminate_ctx_finished_requests()
+
                 # march forward in microbatch slots
                 microbatch_id = (microbatch_id + 1) % self.num_micro_batches
 

tensorrt_llm/_torch/pyexecutor/resource_manager.py

@@ -155,18 +155,33 @@ def __init__(
                 (num_kv_heads + tp_size - 1) // tp_size
                 for _ in range(self.num_local_layers)
             ]
+            self.total_num_kv_heads_per_layer = [
+                (num_kv_heads + tp_size - 1) // tp_size
+                for _ in range(self.num_layers)
+            ]
         else:
             assert len(num_kv_heads) == self.num_layers
 
+            def append_to_kv_heads_per_layer(num_kv_heads_per_layer: List[int],
+                                             kv_head: Optional[int]):
+                if kv_head is not None:
+                    num_kv_heads_per_layer.append(
+                        (kv_head + tp_size - 1) // tp_size)
+                else:
+                    num_kv_heads_per_layer.append(0)
+
             self.num_kv_heads_per_layer = []
             if self.num_local_layers > 0:
                 for i in self.pp_layers:
                     kv_head = num_kv_heads[i]
-                    if kv_head is not None:
-                        self.num_kv_heads_per_layer.append(
-                            (kv_head + tp_size - 1) // tp_size)
-                    else:
-                        self.num_kv_heads_per_layer.append(0)
+                    append_to_kv_heads_per_layer(self.num_kv_heads_per_layer,
+                                                 kv_head)
+
+            self.total_num_kv_heads_per_layer = []
+            for i in range(self.num_layers):
+                kv_head = num_kv_heads[i]
+                append_to_kv_heads_per_layer(self.total_num_kv_heads_per_layer,
+                                             kv_head)
 
         self.num_kv_heads = num_kv_heads
         self.head_dim = head_dim

tests/integration/defs/accuracy/accuracy_core.py

@@ -735,3 +735,14 @@ def setup_class(cls):
         logger.set_level("info")
         yield
         logger.set_level(original_level)
+
+
+def get_accuracy_task(dataset_name: str):
+    try:
+        task_class = globals()[dataset_name]
+        if issubclass(task_class, AccuracyTask):
+            return task_class
+        else:
+            raise ValueError(f"Unknown dataset: {dataset_name}.")
+    except KeyError:
+        raise ValueError(f"Not registered dataset: {dataset_name}.")

tests/integration/defs/accuracy/test_disaggregated_serving.py

@@ -20,9 +20,11 @@
 from tensorrt_llm.llmapi import CompletionOutput, RequestOutput, SamplingParams
 from tensorrt_llm.llmapi.llm_args import LlmArgs
 
-from ..conftest import llm_models_root, parametrize_with_ids, skip_pre_hopper
+from ..conftest import (get_device_count, llm_models_root, parametrize_with_ids,
+                        skip_pre_hopper)
 from ..trt_test_alternative import popen
-from .accuracy_core import GSM8K, MMLU, LlmapiAccuracyTestHarness
+from .accuracy_core import (GSM8K, MMLU, LlmapiAccuracyTestHarness,
+                            get_accuracy_task)
 
 
 class Result(GenerationResultBase):
@@ -71,6 +73,12 @@ def launch_disaggregated_llm(disaggregated_server_config: Dict[str, Any],
     temp_dir = tempfile.TemporaryDirectory()
     disaggregated_serving_config_path = os.path.join(
         temp_dir.name, "disaggregated_serving_config.yaml")
+
+    if tensor_parallel_size > 1:
+        print(
+            f"Using unified tp parameter for testing is not recommended. Please use server configs instead."
+        )
+
     with open(disaggregated_serving_config_path, "w") as f:
         yaml.dump(disaggregated_server_config, f)
     ctx_server_config_path = os.path.join(temp_dir.name,
@@ -88,27 +96,40 @@ def launch_disaggregated_llm(disaggregated_server_config: Dict[str, Any],
     trtllm_serve_path = "trtllm-serve"
     # Common arguments for both servers
     common_args = [
-        trtllm_serve_path, model_name, "--host", "localhost", "--backend",
-        "pytorch"
+        trtllm_serve_path,
+        model_name,
+        "--host",
+        "localhost",
+        "--backend",
+        "pytorch",
     ]
-
-    if tensor_parallel_size > 1:
-        common_args.append(f"--tp_size={tensor_parallel_size}")
+    gen_tp, gen_pp = gen_server_config.get(
+        "tensor_parallel_size",
+        tensor_parallel_size), gen_server_config.get("pipeline_parallel_size",
+                                                     1)
+    ctx_tp, ctx_pp = ctx_server_config.get(
+        "tensor_parallel_size",
+        tensor_parallel_size), ctx_server_config.get("pipeline_parallel_size",
+                                                     1)
+
+    ctx_total_gpus = ctx_tp * ctx_pp
+    gen_total_gpus = gen_tp * gen_pp
 
     env_ctx = os.environ.copy()
     env_ctx["TRTLLM_USE_UCX_KVCACHE"] = "1"
-    env_ctx["CUDA_VISIBLE_DEVICES"] = ",".join(
-        map(str, range(tensor_parallel_size)))
+    env_ctx["CUDA_VISIBLE_DEVICES"] = ",".join(map(str, range(ctx_total_gpus)))
 
     env_gen = os.environ.copy()
     env_gen["TRTLLM_USE_UCX_KVCACHE"] = "1"
     env_gen["CUDA_VISIBLE_DEVICES"] = ",".join(
-        map(str, range(tensor_parallel_size, 2 * tensor_parallel_size)))
+        map(str, range(ctx_total_gpus, ctx_total_gpus + gen_total_gpus)))
     ctx_server_args = common_args + [
-        "--port", "8001", "--extra_llm_api_options", ctx_server_config_path
+        "--port", "8001", "--extra_llm_api_options", ctx_server_config_path,
+        f"--tp_size={ctx_tp}", f"--pp_size={ctx_pp}"
     ]
     gen_server_args = common_args + [
-        "--port", "8002", "--extra_llm_api_options", gen_server_config_path
+        "--port", "8002", "--extra_llm_api_options", gen_server_config_path,
+        f"--tp_size={gen_tp}", f"--pp_size={gen_pp}"
     ]
     if "max_num_tokens" in ctx_server_config:
         ctx_server_args.append(
@@ -182,6 +203,56 @@ def generate_async(prompt: str,
             disaggregated_server.wait()
 
 
+def run_parallel_test(model_name: str, model_path: str, ctx_pp: int,
+                      ctx_tp: int, gen_pp: int, gen_tp: int,
+                      test_set: LlmapiAccuracyTestHarness):
+    if ctx_tp * ctx_pp + gen_tp * gen_pp > get_device_count():
+        pytest.fail(
+            f"Not enough devices for ctx_pp={ctx_pp}+ctx_tp={ctx_tp} and gen_pp={gen_pp}+gen_tp={gen_tp} test"
+        )
+
+    kv_cache_config = {
+        "free_gpu_memory_fraction": 0.5,
+        "enable_block_reuse": False
+    }
+    ctx_server_config = {
+        "pipeline_parallel_size": ctx_pp,
+        "tensor_parallel_size": ctx_tp,
+        "disable_overlap_scheduler": True,
+        "kv_cache_config": kv_cache_config,
+        "cache_transceiver_config": {
+            "backend": "default"
+        }
+    }
+    gen_server_config = {
+        "tensor_parallel_size": gen_tp,
+        "pipeline_parallel_size": gen_pp,
+        "disable_overlap_scheduler": True,
+        "kv_cache_config": kv_cache_config,
+        "cache_transceiver_config": {
+            "backend": "default"
+        }
+    }
+    disaggregated_server_config = {
+        "hostname": "localhost",
+        "port": 8000,
+        "backend": "pytorch",
+        "context_servers": {
+            "num_instances": 1,
+            "urls": ["localhost:8001"]
+        },
+        "generation_servers": {
+            "num_instances": 1,
+            "urls": ["localhost:8002"]
+        }
+    }
+    with launch_disaggregated_llm(disaggregated_server_config,
+                                  ctx_server_config, gen_server_config,
+                                  model_path) as llm:
+        task = test_set(model_name)
+        task.evaluate(llm)
+
+
 @pytest.mark.timeout(3600)
 class TestLlama3_1_8BInstruct(LlmapiAccuracyTestHarness):
     MODEL_NAME = "meta-llama/Llama-3.1-8B-Instruct"
@@ -315,6 +386,20 @@ def test_eagle3(self, overlap_scheduler):
             task = GSM8K(self.MODEL_NAME)
             task.evaluate(llm)
 
+    @pytest.mark.parametrize("tp,pp", [(1, 2), (2, 1), (2, 2)],
+                             ids=["tp1pp2", "tp2pp1", "tp2pp2"])
+    @pytest.mark.parametrize("testset", ["GSM8K", "MMLU"])
+    def test_tp_pp_symmetric(self, tp, pp, testset):
+        return run_parallel_test(self.MODEL_NAME, self.MODEL_PATH, pp, tp, pp,
+                                 tp, get_accuracy_task(testset))
+
+    @parametrize_with_ids("ctx_pp", [2, 4])
+    @parametrize_with_ids("gen_tp", [1, 2])
+    @pytest.mark.parametrize("testset", ["GSM8K", "MMLU"])
+    def test_ctx_pp_gen_tp_asymmetric(self, ctx_pp, gen_tp, testset):
+        return run_parallel_test(self.MODEL_NAME, self.MODEL_PATH, ctx_pp, 1, 1,
+                                 gen_tp, get_accuracy_task(testset))
+
 
 @pytest.mark.skip_less_device_memory(140000)
 @pytest.mark.timeout(3600)