[MetaxGPU] adapt to the latest fastdeploy on metax gpu (#3492)

custom_ops/setup_ops.py

@@ -591,6 +591,12 @@ elif paddle.device.is_compiled_with_custom_device("metax_gpu"):
         if not os.listdir(json_dir):
             raise ValueError("Git clone nlohmann_json failed!")
     sources = [
+        "gpu_ops/update_inputs_v1.cu",
+        "gpu_ops/save_with_output_msg.cc",
+        "gpu_ops/get_output.cc",
+        "gpu_ops/get_output_msg_with_topk.cc",
+        "gpu_ops/save_output_msg_with_topk.cc",
+        "gpu_ops/transfer_output.cc",
         "gpu_ops/save_with_output.cc",
         "gpu_ops/set_mask_value.cu",
         "gpu_ops/set_value_by_flags.cu",

docs/get_started/installation/metax_gpu.md

@@ -0,0 +1,83 @@
+# Metax GPU Installation for running ERNIE 4.5 Series Models
+
+The following installation methods are available when your environment meets these requirements:
+- Python >= 3.10
+- Linux X86_64
+
+Before starting, prepare a machine equipped with Enflame S60 accelerator cards. Requirements:
+
+| Chip Type | Driver Version | KMD Version |
+| :---: | :---: | :---: |
+| MetaX C550 | 3.0.0.1  | 2.14.6 |
+
+## 1. Pre-built Docker Installation (Recommended)
+
+```shell
+docker login --username=cr_temp_user --password=eyJpbnN0YW5jZUlkIjoiY3JpLXpxYTIzejI2YTU5M3R3M2QiLCJ0aW1lIjoiMTc1NTUxODEwODAwMCIsInR5cGUiOiJzdWIiLCJ1c2VySWQiOiIyMDcwOTQwMTA1NjYzNDE3OTIifQ:8226ca50ce5476c42062e24d3c465545de1c1780 cr.metax-tech.com && docker pull cr.metax-tech.com/public-library/maca-native:3.0.0.4-ubuntu20.04-amd64
+```
+
+## 2. paddlepaddle and custom device installation
+
+```shell
+1）pip install paddlepaddle==3.0.0.dev20250729 -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
+2）pip install paddle-metax-gpu==3.0.0.dev20250807 -i https://www.paddlepaddle.org.cn/packages/nightly/maca/
+```
+
+## 3. Build Wheel from Source
+Then clone the source code and build:
+```shell
+git clone https://github.com/PaddlePaddle/FastDeploy
+cd FastDeploy
+bash build.sh
+```
+The built packages will be in the ```FastDeploy/dist``` directory.
+
+## 4. Environment Verification
+
+After installation, verify the environment with this Python code:
+```python
+import paddle
+from paddle.jit.marker import unified
+# Verify GPU availability
+paddle.utils.run_check()
+# Verify FastDeploy custom operators compilation
+from fastdeploy.model_executor.ops.gpu import beam_search_softmax
+```
+
+If the above code executes successfully, the environment is ready.
+
+## 5. Demo
+from fastdeploy import LLM, SamplingParams
+
+prompts = [
+    "Hello. My name is",
+]
+
+sampling_params = SamplingParams(top_p=0.95, max_tokens=32, temperature=0.6)
+
+llm = LLM(model="/root/model/ERNIE-4.5-21B-A3B-Paddle", tensor_parallel_size=1, max_model_len=256, engine_worker_queue_port=9135, quantization='wint8', static_decode_blocks=0, gpu_memory_utilization=0.9)
+
+outputs = llm.generate(prompts, sampling_params)
+
+print(f"Generated {len(outputs)} outputs")
+print("=" * 50 + "\n")
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs.text
+    print(prompt)
+    print(generated_text)
+    print("-" * 50)
+
+Output：
+INFO     2025-08-18 10:54:18,455 416822 engine.py[line:202] Waiting worker processes ready...
+Loading Weights: 100%|█████████████████████████████████████████████████████████████████████████| 100/100 [03:33<00:00,  2.14s/it]
+Loading Layers: 100%|██████████████████████████████████████████████████████████████████████████| 100/100 [00:18<00:00,  5.54it/s]
+INFO     2025-08-18 10:58:16,149 416822 engine.py[line:247] Worker processes are launched with 240.08204197883606 seconds.
+Processed prompts: 100%|███████████████████████| 1/1 [00:21<00:00, 21.84s/it, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+Generated 1 outputs
+==================================================
+
+Hello. My name is
+Alice and I'm here to help you. What can I do for you today?
+Hello Alice! I'm trying to organize a small party

docs/zh/get_started/installation/metax_gpu.md

@@ -0,0 +1,82 @@
+# 使用 Metax GPU C550 运行ERNIE 4.5 系列模型
+
+FastDeploy在Metax C550上对ERNIE 4.5系列模型进行了深度适配和优化，实现了推理入口和GPU的统一，无需修改即可完成推理任务的迁移。
+
+环境准备：
+- Python >= 3.10
+- Linux X86_64
+
+| Chip Type | Driver Version | KMD Version |
+| :---: | :---: | :---: |
+| MetaX C550 | 3.0.0.1  | 2.14.6 |
+
+## 1. 容器镜像获取
+
+```shell
+docker login --username=cr_temp_user --password=eyJpbnN0YW5jZUlkIjoiY3JpLXpxYTIzejI2YTU5M3R3M2QiLCJ0aW1lIjoiMTc1NTUxODEwODAwMCIsInR5cGUiOiJzdWIiLCJ1c2VySWQiOiIyMDcwOTQwMTA1NjYzNDE3OTIifQ:8226ca50ce5476c42062e24d3c465545de1c1780 cr.metax-tech.com && docker pull cr.metax-tech.com/public-library/maca-native:3.0.0.4-ubuntu20.04-amd64
+```
+
+## 2. 预安装
+
+```shell
+1）pip install paddlepaddle==3.0.0.dev20250729 -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
+2）pip install paddle-metax-gpu==3.0.0.dev20250807 -i https://www.paddlepaddle.org.cn/packages/nightly/maca/
+```
+
+## 3. FastDeploy代码下载并编译
+
+```shell
+git clone https://github.com/PaddlePaddle/FastDeploy
+cd FastDeploy
+bash build.sh
+```
+The built packages will be in the ```FastDeploy/dist``` directory.
+
+## 4. 环境验证
+
+After installation, verify the environment with this Python code:
+```python
+import paddle
+from paddle.jit.marker import unified
+# Verify GPU availability
+paddle.utils.run_check()
+# Verify FastDeploy custom operators compilation
+from fastdeploy.model_executor.ops.gpu import beam_search_softmax
+```
+If the above code executes successfully, the environment is ready.
+
+## 5. 示例
+from fastdeploy import LLM, SamplingParams
+
+prompts = [
+    "Hello. My name is",
+]
+
+sampling_params = SamplingParams(top_p=0.95, max_tokens=32, temperature=0.6)
+
+llm = LLM(model="/root/model/ERNIE-4.5-21B-A3B-Paddle", tensor_parallel_size=1, max_model_len=256, engine_worker_queue_port=9135, quantization='wint8', static_decode_blocks=0, gpu_memory_utilization=0.9)
+
+outputs = llm.generate(prompts, sampling_params)
+
+print(f"Generated {len(outputs)} outputs")
+print("=" * 50 + "\n")
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs.text
+    print(prompt)
+    print(generated_text)
+    print("-" * 50)
+
+输出：
+INFO     2025-08-18 10:54:18,455 416822 engine.py[line:202] Waiting worker processes ready...
+Loading Weights: 100%|█████████████████████████████████████████████████████████████████████████| 100/100 [03:33<00:00,  2.14s/it]
+Loading Layers: 100%|██████████████████████████████████████████████████████████████████████████| 100/100 [00:18<00:00,  5.54it/s]
+INFO     2025-08-18 10:58:16,149 416822 engine.py[line:247] Worker processes are launched with 240.08204197883606 seconds.
+Processed prompts: 100%|███████████████████████| 1/1 [00:21<00:00, 21.84s/it, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+Generated 1 outputs
+==================================================
+
+Hello. My name is
+Alice and I'm here to help you. What can I do for you today?
+Hello Alice! I'm trying to organize a small party

fastdeploy/model_executor/layers/backends/metax/attention/flash_attn_backend.py

@@ -257,6 +257,7 @@ class FlashAttentionBackend(AttentionBackend):
         out = paddle.add(paddle.multiply(qk, cos), paddle.multiply(rotate_half, sin))
         return paddle.cast(out, qk.dtype)
 
+    @paddle.no_grad()
     def forward_native_backend(
         self,
         q: paddle.Tensor,
@@ -273,7 +274,7 @@ class FlashAttentionBackend(AttentionBackend):
         # 1. 分离 encoder / decoder 的 mask
         seq_lens_encoder = forward_meta.seq_lens_encoder.squeeze(-1)
         seq_lens_decoder = forward_meta.seq_lens_decoder.squeeze(-1)
-        seq_lens_this_time = forward_meta.seq_lens_this_time.squeeze(-1)
+        seq_lens_this_time = forward_meta.seq_lens_this_time
         encoder_indices = []
         decoder_indices = []
 

fastdeploy/model_executor/layers/backends/metax/moe/fused_moe_triton_metax_backend.py

@@ -45,6 +45,7 @@ class MetaxTritonWeightOnlyMoEMethod(QuantMethodBase):
         """process_prequanted_weights"""
         pass
 
+    @paddle.no_grad()
     def create_weights(self, layer: nn.Layer, state_dict):
         """
         Triton MoE create weight process.
@@ -125,11 +126,12 @@ class MetaxTritonWeightOnlyMoEMethod(QuantMethodBase):
                 )
                 getattr(layer, scale_name).set_value(quanted_weight_scale)
 
+    @paddle.no_grad()
     def apply(
         self,
         layer: nn.Layer,
         x: paddle.Tensor,
-        gate_out: paddle.Tensor,
+        gate: nn.Layer,
     ) -> paddle.Tensor:
         """
         Triton compute Fused MoE.
@@ -141,6 +143,7 @@ class MetaxTritonWeightOnlyMoEMethod(QuantMethodBase):
         moe_intermediate_size = layer.moe_intermediate_size
         hidden_size = layer.hidden_size
 
+        gate_out = gate(x.cast("float32"))
         topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
             gate_out,
             layer.gate_correction_bias,

fastdeploy/model_executor/pre_and_post_process.py

@@ -52,6 +52,7 @@ elif current_platform.is_maca():
         set_stop_value_multi_ends,
         step_paddle,
         update_inputs,
+        update_inputs_v1,
     )
 else:
     from fastdeploy.model_executor.ops.gpu import (

fastdeploy/worker/metax_model_runner.py

@@ -23,8 +23,11 @@ import paddle
 from paddle import nn
 from paddleformers.utils.log import logger
 
+from fastdeploy import envs
 from fastdeploy.config import FDConfig
 from fastdeploy.engine.request import Request, RequestType
+from fastdeploy.input.mm_processor import DataProcessor
+from fastdeploy.model_executor.forward_meta import ForwardMeta
 from fastdeploy.model_executor.graph_optimization.utils import (
     profile_run_guard,
     sot_warmup_guard,
@@ -41,6 +44,7 @@ from fastdeploy.model_executor.layers.rotary_embedding import get_rope, get_rope
 from fastdeploy.model_executor.layers.sample.meta_data import SamplingMetadata
 from fastdeploy.model_executor.layers.sample.sampler import Sampler, SpeculativeSampler
 from fastdeploy.model_executor.model_loader import get_model_loader
+from fastdeploy.model_executor.models.ernie4_5_vl.modeling_resampler import ScatterOp
 from fastdeploy.model_executor.ops.gpu import (
     recover_decode_task,
     set_value_by_flags_and_idx,
@@ -52,15 +56,7 @@ from fastdeploy.model_executor.pre_and_post_process import (
     rebuild_padding,
     step_cuda,
 )
-from fastdeploy.platforms import current_platform
+from fastdeploy.spec_decode import MTPProposer, NgramProposer
-
-if not current_platform.is_dcu():
-    from fastdeploy.spec_decode import MTPProposer, NgramProposer
-
-from fastdeploy import envs
-from fastdeploy.input.mm_processor import DataProcessor
-from fastdeploy.model_executor.forward_meta import ForwardMeta
-from fastdeploy.model_executor.models.ernie4_5_vl.modeling_resampler import ScatterOp
 from fastdeploy.worker.model_runner_base import ModelRunnerBase
 from fastdeploy.worker.output import ModelOutputData, ModelRunnerOutput
 
@@ -130,7 +126,7 @@ class MetaxModelRunner(ModelRunnerBase):
             shape=[self.parallel_config.max_num_seqs, 1],
             fill_value=4,
             dtype="int64",
-        )
+        ).cpu()
         self.restore_chunked_prefill_request = dict()
 
         # Initialize attention Backend
@@ -164,6 +160,7 @@ class MetaxModelRunner(ModelRunnerBase):
         if self.speculative_method == "ngram":
             self.proposer = NgramProposer(self.fd_config)
         elif self.speculative_method == "mtp":
+            self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer
             self.proposer = MTPProposer(
                 self.fd_config,
                 self.get_model(),
@@ -193,21 +190,23 @@ class MetaxModelRunner(ModelRunnerBase):
 
         return self.guided_backend.get_logits_processor(schemata_key=schemata_key), schemata_key
 
-    def insert_tasks_v1(self, req_dicts: List[Request]):
+    def insert_tasks_v1(self, req_dicts: List[Request], num_running_requests: int = None):
         """
         Process scheduler output tasks, used when ENABLE_V1_KVCACHE_SCHEDULER=1
+        req_dict: A list of Request dict
+        num_running_requests: batch_size
         """
-        # NOTE(luotingdan): Lazy initialize kv cache
+        # Lazy initialize kv cache
         if "caches" not in self.share_inputs:
             self.initialize_kv_cache()
 
         req_len = len(req_dicts)
         has_prefill_task = False
+        has_decode_task = False
         for i in range(req_len):
             request = req_dicts[i]
             idx = request.idx
             if request.task_type.value == RequestType.PREFILL.value:  # prefill task
-                logger.debug(f"Handle prefill request {request} at idx {idx}")
                 prefill_start_index = request.prefill_start_index
                 prefill_end_index = request.prefill_end_index
                 length = prefill_end_index - prefill_start_index
@@ -253,6 +252,11 @@ class MetaxModelRunner(ModelRunnerBase):
                     )
 
                 input_ids = request.prompt_token_ids + request.output_token_ids
+                logger.debug(
+                    f"Handle prefill request {request} at idx {idx}, "
+                    f"{prefill_start_index=}, {prefill_end_index=}, "
+                    f"need_prefilled_token_num={len(input_ids)}"
+                )
                 self.share_inputs["input_ids"][idx : idx + 1, :length] = np.array(
                     input_ids[prefill_start_index:prefill_end_index]
                 )
@@ -264,7 +268,7 @@ class MetaxModelRunner(ModelRunnerBase):
                 )
                 self.share_inputs["stop_flags"][idx : idx + 1] = False
                 self.share_inputs["seq_lens_decoder"][idx : idx + 1] = prefill_start_index
-                self.share_inputs["seq_lens_this_time"][idx : idx + 1] = length
+                self.seq_lens_this_time_buffer[idx : idx + 1] = length
                 self.share_inputs["seq_lens_encoder"][idx : idx + 1] = length
                 self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = 0
                 self.share_inputs["prompt_lens"][idx : idx + 1] = len(input_ids)
@@ -281,22 +285,27 @@ class MetaxModelRunner(ModelRunnerBase):
                 self.share_inputs["block_tables"][idx : idx + 1, :encoder_block_num] = np.array(
                     request.block_tables, dtype="int32"
                 )
+                if self.share_inputs["is_block_step"][idx]:  # has tasks to continue to decode
+                    has_decode_task = True
                 continue
             else:  # preempted task
                 logger.debug(f"Handle preempted request {request} at idx {idx}")
                 self.share_inputs["block_tables"][idx : idx + 1, :] = -1
                 self.share_inputs["stop_flags"][idx : idx + 1] = True
-                self.share_inputs["seq_lens_this_time"][idx : idx + 1] = 0
+                self.seq_lens_this_time_buffer[idx : idx + 1] = 0
                 self.share_inputs["seq_lens_decoder"][idx : idx + 1] = 0
                 self.share_inputs["seq_lens_encoder"][idx : idx + 1] = 0
                 self.share_inputs["is_block_step"][idx : idx + 1] = False
                 continue
 
-            if len(request.eos_token_ids) < self.parallel_config.eos_tokens_lens:
+            assert len(request.eos_token_ids) == self.model_config.eos_tokens_lens
-                request.eos_token_ids.append(request.eos_token_ids[0])
             self.share_inputs["eos_token_id"][:] = np.array(request.eos_token_ids, dtype="int64").reshape(-1, 1)
 
             self.share_inputs["top_p"][idx : idx + 1] = request.get("top_p", 0.7)
+            self.share_inputs["top_k"][idx : idx + 1] = request.get("top_k", 0)
+            self.share_inputs["top_k_list"][idx] = request.get("top_k", 0)
+            self.share_inputs["min_p"][idx : idx + 1] = request.get("min_p", 0.0)
+            self.share_inputs["min_p_list"][idx] = request.get("min_p", 0.0)
             self.share_inputs["temperature"][idx : idx + 1] = request.get("temperature", 0.95)
             self.share_inputs["penalty_score"][idx : idx + 1] = request.get("repetition_penalty", 1.0)
             self.share_inputs["frequency_score"][idx : idx + 1] = request.get("frequency_penalty", 0.0)
@@ -326,12 +335,15 @@ class MetaxModelRunner(ModelRunnerBase):
             else:
                 self.share_inputs["stop_seqs_len"][idx : idx + 1, :] = 0
 
-        if has_prefill_task:
+        if has_prefill_task or has_decode_task:
             self.share_inputs["not_need_stop"][0] = True
+        self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer[:num_running_requests]
 
-    def insert_prefill_inputs(self, req_dicts: List[Request]):
+    def insert_prefill_inputs(self, req_dicts: List[Request], num_running_requests: int = None):
         """
         Process inputs for prefill tasks and insert it to share_inputs buffer
+        req_dict: A list of Request dict
+        num_running_requests: batch_size
         TODO(gongshaotian): Refactor this func
         """
 
@@ -365,7 +377,7 @@ class MetaxModelRunner(ModelRunnerBase):
                 self.share_inputs["prompt_ids"][idx : idx + 1, :length] = np.array(request.prompt_token_ids)
                 self.share_inputs["seq_lens_encoder"][idx : idx + 1] = 0
                 self.share_inputs["seq_lens_decoder"][idx : idx + 1] = length
-                self.share_inputs["seq_lens_this_time"][idx : idx + 1] = 1
+                self.seq_lens_this_time_buffer[idx : idx + 1] = 1
                 self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = 0
                 self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = length
                 self.share_inputs["prompt_lens"][idx : idx + 1] = length
@@ -377,7 +389,7 @@ class MetaxModelRunner(ModelRunnerBase):
                         request.draft_token_ids[0:num_prefill_send_token],
                         dtype="int64",
                     )
-                    self.share_inputs["seq_lens_this_time"][idx : idx + 1] = num_prefill_send_token
+                    self.seq_lens_this_time_buffer[idx : idx + 1] = num_prefill_send_token
             else:
                 self.share_inputs["pre_ids"][idx : idx + 1] = -1
                 self.share_inputs["step_idx"][idx : idx + 1] = 0
@@ -412,7 +424,7 @@ class MetaxModelRunner(ModelRunnerBase):
                         )
                         self.share_inputs["seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
                         self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
-                    self.share_inputs["seq_lens_this_time"][idx : idx + 1] = token_chunk_size
+                    self.seq_lens_this_time_buffer[idx : idx + 1] = token_chunk_size
                     self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = token_chunk_size
                     self.share_inputs["seq_lens_encoder"][idx : idx + 1] = token_chunk_size
                     self.share_inputs["prompt_lens"][idx : idx + 1] = token_chunk_size
@@ -430,7 +442,7 @@ class MetaxModelRunner(ModelRunnerBase):
                     else:
                         self.share_inputs["seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
                         self.share_inputs["step_seq_lens_decoder"][idx : idx + 1] = request.get("seq_lens_decoder", 0)
-                    self.share_inputs["seq_lens_this_time"][idx : idx + 1] = length
+                    self.seq_lens_this_time_buffer[idx : idx + 1] = length
                     self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = length
                     self.share_inputs["seq_lens_encoder"][idx : idx + 1] = length
                     self.share_inputs["prompt_lens"][idx : idx + 1] = length
@@ -453,12 +465,13 @@ class MetaxModelRunner(ModelRunnerBase):
                 else:
                     return default_value
 
-            if len(request.eos_token_ids) < self.parallel_config.eos_tokens_lens:
+            assert len(request.eos_token_ids) == self.model_config.eos_tokens_lens
-                request.eos_token_ids.append(request.eos_token_ids[0])
             self.share_inputs["eos_token_id"][:] = np.array(request.eos_token_ids, dtype="int64").reshape(-1, 1)
             self.share_inputs["top_p"][idx : idx + 1] = get_attr_from_request(request, "top_p", 0.7)
             self.share_inputs["top_k"][idx : idx + 1] = request.get("top_k", 0)
+            self.share_inputs["top_k_list"][idx] = request.get("top_k", 0)
             self.share_inputs["min_p"][idx : idx + 1] = request.get("min_p", 0.0)
+            self.share_inputs["min_p_list"][idx] = request.get("min_p", 0.0)
 
             self.share_inputs["temperature"][idx : idx + 1] = get_attr_from_request(request, "temperature", 0.95)
             self.share_inputs["penalty_score"][idx : idx + 1] = get_attr_from_request(
@@ -489,13 +502,15 @@ class MetaxModelRunner(ModelRunnerBase):
                 request.block_tables, dtype="int32"
             )
 
-            if request.get("bad_words_token_ids") is not None:
+            if request.get("bad_words_token_ids") is not None and len(request.get("bad_words_token_ids")) > 0:
                 bad_words_len = len(request.get("bad_words_token_ids"))
-                if bad_words_len > 0:
                 self.share_inputs["bad_tokens_len"][idx : idx + 1] = bad_words_len
                 self.share_inputs["bad_tokens"][idx : idx + 1, :bad_words_len] = np.array(
                     request.get("bad_words_token_ids"), dtype="int64"
                 )
+            else:
+                self.share_inputs["bad_tokens_len"][idx : idx + 1] = 1
+                self.share_inputs["bad_tokens"][idx : idx + 1, :] = np.array([-1], dtype="int64")
 
             if request.get("stop_token_ids") is not None and request.get("stop_seqs_len") is not None:
                 stop_seqs_num = len(request.get("stop_seqs_len"))
@@ -514,8 +529,10 @@ class MetaxModelRunner(ModelRunnerBase):
 
         self.share_inputs["not_need_stop"][0] = True
 
+        self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer[:num_running_requests]
+
         if self.speculative_method in ["mtp"]:
-            self.proposer.insert_prefill_inputs(req_dicts)
+            self.proposer.insert_prefill_inputs(req_dicts, num_running_requests)
 
     def _dummy_prefill_inputs(self, num_tokens: int, batch_size: int, expected_decode_len: int):
         """Set dummy prefill inputs to share_inputs"""
@@ -525,6 +542,12 @@ class MetaxModelRunner(ModelRunnerBase):
             num_tokens // batch_size,
             self.parallel_config.max_model_len - max_dec_len,
         )
+
+        # When the full length is too large, DeepEP's buffer size will not be enough to cause the result to appear nan.
+        # Figure out the accurate buffer size of DeepEP.
+        if self.fd_config.parallel_config.enable_expert_parallel:
+            full_length = min(full_length, 32)
+
         input_length = int(full_length * self.cache_config.kv_cache_ratio)
         block_num = (
             input_length + self.cache_config.block_size - 1
@@ -534,8 +557,10 @@ class MetaxModelRunner(ModelRunnerBase):
             idx = i
             self.share_inputs["input_ids"][idx : idx + 1, :input_length] = np.array([5] * input_length)
             self.share_inputs["prompt_ids"][idx : idx + 1, :input_length] = np.array([5] * input_length)
-            self.share_inputs["eos_token_id"][:] = np.array([2], dtype="int64").reshape(-1, 1)
+            self.share_inputs["eos_token_id"][:] = np.array(
-            self.share_inputs["seq_lens_this_time"][idx : idx + 1] = input_length
+                [2] * self.model_config.eos_tokens_lens, dtype="int64"
+            ).reshape(-1, 1)
+            self.seq_lens_this_time_buffer[idx : idx + 1] = input_length
             self.share_inputs["step_seq_lens_encoder"][idx : idx + 1] = input_length
             self.share_inputs["seq_lens_encoder"][idx : idx + 1] = input_length
             self.share_inputs["seq_lens_decoder"][idx : idx + 1] = 0
@@ -553,6 +578,7 @@ class MetaxModelRunner(ModelRunnerBase):
             self.share_inputs["block_tables"][idx : idx + 1, :block_num] = np.arange(
                 idx * block_num, (idx + 1) * block_num, 1
             )
+        self.share_inputs["seq_lens_this_time"] = self.seq_lens_this_time_buffer
 
     def _init_share_inputs(self, max_num_seqs: int):
         """
@@ -568,18 +594,20 @@ class MetaxModelRunner(ModelRunnerBase):
         )
         self.share_inputs["input_ids"] = paddle.full(
             [max_num_seqs, self.parallel_config.max_model_len],
-            self.parallel_config.pad_token_id,
+            self.model_config.pad_token_id,
             dtype="int64",
         )
         self.share_inputs["prompt_ids"] = paddle.full(
             [max_num_seqs, self.parallel_config.max_model_len],
-            self.parallel_config.pad_token_id,
+            self.model_config.pad_token_id,
             dtype="int64",
         )
-        self.share_inputs["eos_token_id"] = paddle.full([self.parallel_config.eos_tokens_lens, 1], 0, dtype="int64")
+        self.share_inputs["eos_token_id"] = paddle.full([self.model_config.eos_tokens_lens, 1], 0, dtype="int64")
         self.share_inputs["top_p"] = paddle.full([max_num_seqs, 1], self.model_config.top_p, dtype="float32")
         self.share_inputs["top_k"] = paddle.full([max_num_seqs, 1], 0, dtype="int64")
+        self.share_inputs["top_k_list"] = [0] * max_num_seqs
         self.share_inputs["min_p"] = paddle.full([max_num_seqs, 1], 0.0, dtype="float32")
+        self.share_inputs["min_p_list"] = [0.0] * max_num_seqs
         self.share_inputs["temperature"] = paddle.full(
             [max_num_seqs, 1], self.model_config.temperature, dtype="float32"
         )
@@ -603,7 +631,9 @@ class MetaxModelRunner(ModelRunnerBase):
         self.share_inputs["max_length"] = paddle.full(
             [max_num_seqs, 1], self.model_config.max_model_len, dtype="int64"
         )
-        self.share_inputs["seq_lens_this_time"] = paddle.full(max_num_seqs, 0, dtype="int32")
+        self.seq_lens_this_time_buffer = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+        if self.fd_config.parallel_config.enable_expert_parallel:
+            self.share_inputs["seq_lens_this_time"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["seq_lens_decoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["step_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
@@ -626,7 +656,7 @@ class MetaxModelRunner(ModelRunnerBase):
         self.share_inputs["need_block_list"] = paddle.full([max_num_seqs], -1, dtype="int32")
         self.share_inputs["need_block_len"] = paddle.full([1], 0, dtype="int32")
         self.share_inputs["used_list_len"] = paddle.full([max_num_seqs], 0, dtype="int32")
-        self.share_inputs["infer_seed"] = paddle.full([max_num_seqs, 1], 0, dtype="int64")
+        self.share_inputs["infer_seed"] = paddle.full([max_num_seqs, 1], 0, dtype="int64").cpu()
         self.share_inputs["first_token_ids"] = paddle.full([max_num_seqs, 1], -1, dtype="int64")
         self.share_inputs["ori_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
         self.share_inputs["system_lens"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
@@ -637,10 +667,11 @@ class MetaxModelRunner(ModelRunnerBase):
             0,
             dtype="int64",
         )
-        self.share_inputs["cum_offsets"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+        self.share_inputs["batch_id_per_token"] = paddle.full(
-        self.share_inputs["batch_id_per_token"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+            [max_num_seqs * self.parallel_config.max_model_len, 1], 0, dtype="int32"
-        self.share_inputs["cu_seqlens_q"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+        )
-        self.share_inputs["cu_seqlens_k"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
+        self.share_inputs["cu_seqlens_q"] = paddle.full([max_num_seqs + 1, 1], 0, dtype="int32")
+        self.share_inputs["cu_seqlens_k"] = paddle.full([max_num_seqs + 1, 1], 0, dtype="int32")
 
         # Declare AttentionBackend buffers
         self.share_inputs["decoder_batch_ids"] = None
@@ -758,7 +789,6 @@ class MetaxModelRunner(ModelRunnerBase):
         # Remove padding
         (
             ids_remove_padding,
-            cum_offsets,
             batch_id_per_token,
             cu_seqlens_q,
             cu_seqlens_k,
@@ -774,7 +804,6 @@ class MetaxModelRunner(ModelRunnerBase):
         )
 
         self.share_inputs["ids_remove_padding"].copy_(ids_remove_padding, False)
-        self.share_inputs["cum_offsets"].copy_(cum_offsets, False)
         self.share_inputs["batch_id_per_token"].copy_(batch_id_per_token, False)
         self.share_inputs["cu_seqlens_q"].copy_(cu_seqlens_q, False)
         self.share_inputs["cu_seqlens_k"].copy_(cu_seqlens_k, False)
@@ -795,7 +824,10 @@ class MetaxModelRunner(ModelRunnerBase):
             temperature=self.share_inputs["temperature"],
             top_p=self.share_inputs["top_p"],
             top_k=self.share_inputs["top_k"],
+            top_k_list=self.share_inputs["top_k_list"],
             min_p=self.share_inputs["min_p"],
+            min_p_list=self.share_inputs["min_p_list"],
+            seed=self.share_inputs["infer_seed"],
             step_idx=self.share_inputs["step_idx"],
             pre_token_ids=self.share_inputs["pre_ids"],
             prompt_ids=self.share_inputs["prompt_ids"],
@@ -933,7 +965,7 @@ class MetaxModelRunner(ModelRunnerBase):
             self.share_inputs["caches"] = list(cache_kvs.values())
             for value in cache_kvs.values():
                 del value
-        paddle.device.cuda.empty_cache()
+        # paddle.device.empty_cache()
 
     def initialize_attn_backend(self) -> None:
         """
@@ -1023,7 +1055,7 @@ class MetaxModelRunner(ModelRunnerBase):
 
                 hidden_states = rebuild_padding(
                     model_output,
-                    self.share_inputs["cum_offsets"],
+                    self.share_inputs["cu_seqlens_q"],
                     self.share_inputs["seq_lens_this_time"],
                     self.share_inputs["seq_lens_decoder"],
                     self.share_inputs["seq_lens_encoder"],
@@ -1247,6 +1279,7 @@ class MetaxModelRunner(ModelRunnerBase):
     def execute_model(
         self,
         model_forward_batch: Optional[List[Request]] = None,
+        num_running_requests: int = None,
     ) -> Optional[ModelRunnerOutput]:
         """
         The Entrance of model execute.
@@ -1255,6 +1288,7 @@ class MetaxModelRunner(ModelRunnerBase):
             class at the server level, which is too granular for ModelRunner.
             We plan to replace it with 'ModelForwardBatch'.
             intermediate_tensors:
+            num_running_requests: batch_size
         """
         # 1. Prepare inputs of model and sampler.
         skip_idx_list = self._get_skip_idx(model_forward_batch)
@@ -1286,7 +1320,7 @@ class MetaxModelRunner(ModelRunnerBase):
             )
             hidden_states = rebuild_padding(
                 model_output,
-                self.share_inputs["cum_offsets"],
+                self.share_inputs["cu_seqlens_q"],
                 self.share_inputs["seq_lens_this_time"],
                 self.share_inputs["seq_lens_decoder"],
                 self.share_inputs["seq_lens_encoder"],
@@ -1356,8 +1390,8 @@ class MetaxModelRunner(ModelRunnerBase):
             accept_num=(self.share_inputs["accept_num"] if self.speculative_decoding else None),
             enable_thinking=(self.share_inputs["enable_thinking"] if self.enable_mm else None),
             think_end_id=(self.model_config.think_end_id if self.enable_mm else -1),
-            need_think_end=(self.share_inputs["need_think_end"] if self.enable_mm else None),
+            need_think_end=(self.share_inputs["need_think_end"][:num_running_requests] if self.enable_mm else None),
-            reasoning_index=(self.share_inputs["reasoning_index"] if self.enable_mm else None),
+            reasoning_index=(self.share_inputs["reasoning_index"][:num_running_requests] if self.enable_mm else None),
             stop_token_ids=self.share_inputs["stop_seqs"],
             stop_seqs_len=self.share_inputs["stop_seqs_len"],
         )
@@ -1397,6 +1431,9 @@ class MetaxModelRunner(ModelRunnerBase):
 
             self._update_chunked_prefill(model_forward_batch)
             self._add_cache(model_forward_batch)
+        self.seq_lens_this_time_buffer[:num_running_requests].copy_(
+            self.share_inputs["seq_lens_this_time"][:num_running_requests], False
+        )
         return None
 
     def _add_cache(self, model_forward_batch) -> None:
@@ -1528,7 +1565,7 @@ class MetaxModelRunner(ModelRunnerBase):
         """ " Dynamic model loader use to clear parameters use for RL"""
         self.dynamic_weight_manager.clear_parameters(pid)
         self.clear_cache()
-        paddle.device.cuda.empty_cache()
+        # paddle.device.empty_cache()
         self.dynamic_weight_manager._log_memory("dynamic weight manager clear all memory")
 
     def update_parameters(self, pid):

fastdeploy/worker/metax_worker.py

@@ -20,6 +20,7 @@ import time
 from typing import List, Optional
 
 import paddle
+import pymxsml
 from paddle import nn
 
 from fastdeploy import envs
@@ -60,7 +61,7 @@ class MetaxWorker(WorkerBase):
             paddle.set_default_dtype(self.parallel_config.dtype)
 
             gc.collect()
-            paddle.device.cuda.empty_cache()
+
         else:
             raise RuntimeError(f"Not support device type: {self.device_config.device}")
 
@@ -92,8 +93,12 @@ class MetaxWorker(WorkerBase):
             You may limit the usage of GPU memory
             by adjusting the `gpu_memory_utilization` parameter.
         """
-        """Will implement later"""
 
+        # temporary fix kvcache size to test
+        fd_kvache_mem = os.getenv("FD_METAX_KVCACHE_MEM")
+        if fd_kvache_mem is not None:
+            return int(float(fd_kvache_mem) * 1024**3)
+        else:
             # 1. Record memory state before profile run
             start_time = time.perf_counter()
             Gb = 1024**3
@@ -110,8 +115,6 @@ class MetaxWorker(WorkerBase):
             if os.getenv("MACA_VISIBLE_DEVICES") is not None:
                 device_id = int(os.getenv("MACA_VISIBLE_DEVICES").split(",")[device_id])
 
-        import pymxsml
-
             pymxsml.mxSmlInit()
             info = pymxsml.mxSmlGetMemoryInfo(device_id)
             before_run_meminfo_total = info.vramTotal * 1024
@@ -144,10 +147,8 @@ class MetaxWorker(WorkerBase):
             after_run_meminfo_free = after_run_meminfo_total - after_run_meminfo_used
 
             available_kv_cache_memory = (
-            after_run_meminfo_total * self.cache_config.gpu_memory_utilization
+                after_run_meminfo_free - paddle_peak_increase
-            - after_run_meminfo_used
+            ) * self.cache_config.gpu_memory_utilization
-            - paddle_peak_increase
-        )
             available_kv_cache_memory += model_block_memory_used * self.parallel_config.total_block_num
 
             end_time = time.perf_counter()
@@ -180,19 +181,28 @@ class MetaxWorker(WorkerBase):
     def execute_model(
         self,
         model_forward_batch: Optional[List[Request]] = None,
+        num_running_request: int = None,
     ) -> Optional[ModelRunnerOutput]:
         """ """
-        output = self.model_runner.execute_model(model_forward_batch)
+        output = self.model_runner.execute_model(model_forward_batch, num_running_request)
         return output
 
-    def preprocess_new_task(self, req_dicts: List[Request]) -> None:
+    def preprocess_new_task(self, req_dicts: List[Request], num_running_requests: int) -> None:
         """Process new requests and then start the decode loop
         and workers and modelrunners should not perceive it.
         """
         if envs.ENABLE_V1_KVCACHE_SCHEDULER:
-            self.model_runner.insert_tasks_v1(req_dicts=req_dicts)
+            self.model_runner.insert_tasks_v1(req_dicts=req_dicts, num_running_requests=num_running_requests)
         else:
-            self.model_runner.insert_prefill_inputs(req_dicts=req_dicts)
+            self.model_runner.insert_prefill_inputs(req_dicts=req_dicts, num_running_requests=num_running_requests)
+
+    def graph_optimize_and_warm_up_model(self) -> None:
+        """
+        Perform the warm-up and the graph optimization
+        """
+        if self.model_runner.graph_opt_level >= 1:
+            self.model_runner.sot_warmup()
+        # Todo Triger cuda grpah capture.
 
     def check_health(self) -> bool:
         """ """