Merge pull request #9 from cheunglei/tp_dev

nanovllm/config.py

@@ -9,6 +9,7 @@ class Config:
     max_num_seqs: int = 512
     max_model_len: int = 4096
     gpu_memory_utilization: float = 0.9
+    tensor_parallel_size: int = 1
     enforce_eager: bool = False
     hf_config: AutoConfig | None = None
     eos: int = -1
@@ -17,4 +18,4 @@ class Config:
 
     def __post_init__(self):
         assert self.model
-        assert self.kvcache_block_size % 256 == 0
+        assert self.kvcache_block_size % 256 == 0

nanovllm/engine/llm_engine.py

@@ -1,6 +1,8 @@
+import atexit
 from time import perf_counter
 from tqdm.auto import tqdm
 from transformers import AutoConfig, AutoTokenizer
+import torch.multiprocessing as mp
 
 from nanovllm.config import Config
 from nanovllm.sampling_params import SamplingParams
@@ -19,10 +21,24 @@ class LLMEngine:
         Sequence.block_size = config.kvcache_block_size
         config.hf_config = AutoConfig.from_pretrained(config.model)
         config.max_model_len = min(config.max_model_len, config.hf_config.max_position_embeddings)
+        self.ps = []
+        self.events = []
+        for i in range(1, config.tensor_parallel_size):
+            event = mp.Event()
+            process = mp.Process(target=ModelRunner, args=(config, i, event))
+            process.start()
+            self.ps.append(process)
+            self.events.append(event)
+        self.model_runner = ModelRunner(config, 0, self.events)
         self.tokenizer = AutoTokenizer.from_pretrained(config.model, use_fast=True)
         config.eos = self.tokenizer.eos_token_id
-        self.model_runner = ModelRunner(config)
         self.scheduler = Scheduler(config)
+        atexit.register(self.exit)
+
+    def exit(self):
+        self.model_runner.call("exit")
+        for p in self.ps:
+            p.join()
 
     def add_request(self, prompt: str | list[int], sampling_params: SamplingParams):
         if isinstance(prompt, str):
@@ -32,7 +48,7 @@ class LLMEngine:
 
     def step(self):
         seqs, is_prefill = self.scheduler.schedule()
-        token_ids = self.model_runner.run(seqs, is_prefill)
+        token_ids = self.model_runner.call("run", seqs, is_prefill)
         self.scheduler.postprocess(seqs, token_ids)
         outputs = [(seq.seq_id, seq.completion_token_ids) for seq in seqs if seq.is_finished]
         num_tokens = sum(len(seq) for seq in seqs) if is_prefill else -len(seqs)

nanovllm/engine/model_runner.py

@@ -1,4 +1,8 @@
+import pickle
 import torch
+import torch.distributed as dist
+from multiprocess.synchronize import Event
+from multiprocess.shared_memory import SharedMemory
 
 from nanovllm.config import Config
 from nanovllm.engine.sequence import Sequence
@@ -11,12 +15,17 @@ from nanovllm.utils.loader import load_model
 
 class ModelRunner:
 
-    def __init__(self, config: Config):
+    def __init__(self, config: Config, rank: int, event: Event | list[Event]):
         self.config = config
         hf_config = config.hf_config
         self.block_size = config.kvcache_block_size
         self.enforce_eager = config.enforce_eager
+        self.world_size = config.tensor_parallel_size
+        self.rank = rank
+        self.event = event
     
+        dist.init_process_group("nccl", "tcp://localhost:2333", world_size=self.world_size, rank=rank)
+        torch.cuda.set_device(rank)
         default_dtype = torch.get_default_dtype()
         torch.set_default_dtype(hf_config.torch_dtype)
         torch.set_default_device("cuda")
@@ -29,14 +38,66 @@ class ModelRunner:
         torch.set_default_device("cpu")
         torch.set_default_dtype(default_dtype)
 
+        if self.world_size > 1:
+            if rank == 0:
+                self.shm = SharedMemory(name="nanovllm", create=True, size=2**20)
+                dist.barrier()
+            else:
+                dist.barrier()
+                self.shm = SharedMemory(name="nanovllm")
+                self.loop()
+
+    def exit(self):
+        if self.world_size > 1:
+            self.shm.close()
+            if self.rank == 0:
+                self.shm.unlink()
+        # dist.destroy_process_group()
+
+    def loop(self):
+        while True:
+            method_name, args = self.read_shm()
+            method = getattr(self, method_name, None)
+            assert callable(method)
+            method(*args)
+            if method_name == "exit":
+                break
+
+    def read_shm(self):
+        assert self.world_size > 1 and self.rank
+        self.event.wait()
+        n = int.from_bytes(self.shm.buf[0:4], "little")
+        method_name, *args = pickle.loads(self.shm.buf[4:n+4])
+        self.event.clear()
+        return method_name, args
+
+    def write_shm(self, method_name, *args):
+        assert self.world_size > 1 and not self.rank
+        data = pickle.dumps([method_name, *args])
+        n = len(data)
+        assert n + 4 <= self.shm.size
+        self.shm.buf[0:4] = n.to_bytes(4, "little")
+        self.shm.buf[4:n+4] = data
+        for event in self.event:
+            event.set()
+
+    def call(self, method_name, *args):
+        assert self.rank == 0
+        if self.world_size > 1:
+            self.write_shm(method_name, *args)
+        method = getattr(self, method_name, None)
+        assert callable(method)
+        return method(*args)
+
     def allocate_kv_cache(self, gpu_memory_utilization):
         config = self.config
         hf_config = config.hf_config
         total, used, _ = get_gpu_memory()
         free = total * gpu_memory_utilization - used
-        block_bytes = 2 * hf_config.num_hidden_layers * self.block_size * hf_config.num_key_value_heads * hf_config.head_dim * hf_config.torch_dtype.itemsize
+        num_kv_heads = hf_config.num_key_value_heads // dist.get_world_size()
+        block_bytes = 2 * hf_config.num_hidden_layers * self.block_size * num_kv_heads * hf_config.head_dim * hf_config.torch_dtype.itemsize
         config.num_kvcache_blocks = int(free) // block_bytes
-        self.kv_cache = torch.zeros(2, hf_config.num_hidden_layers, config.num_kvcache_blocks, self.block_size, hf_config.num_key_value_heads, hf_config.head_dim)
+        self.kv_cache = torch.zeros(2, hf_config.num_hidden_layers, config.num_kvcache_blocks, self.block_size, num_kv_heads, hf_config.head_dim)
         layer_id = 0
         for module in self.model.modules():
             if hasattr(module, "k_cache") and hasattr(module, "v_cache"):
@@ -148,7 +209,7 @@ class ModelRunner:
         input_ids, positions = self.prepare_prefill(seqs) if is_prefill else self.prepare_decode(seqs)
         temperatures = self.prepare_sample(seqs)
         logits = self.run_model(input_ids, positions, is_prefill)
-        token_ids = self.sampler(logits, temperatures).tolist()
+        token_ids = self.sampler(logits, temperatures).tolist() if self.rank == 0 else None
         reset_context()
         return token_ids
 

nanovllm/engine/scheduler.py

@@ -14,8 +14,6 @@ class Scheduler:
         self.block_manager = BlockManager(config.num_kvcache_blocks, config.kvcache_block_size)
         self.waiting: deque[Sequence] = deque()
         self.running: deque[Sequence] = deque()
-        self.num_finished = 0
-        self.num_tokens = 0
 
     def is_finished(self):
         return not self.waiting and not self.running
@@ -67,11 +65,9 @@ class Scheduler:
         self.waiting.appendleft(seq)
 
     def postprocess(self, seqs: list[Sequence], token_ids: list[int]) -> list[bool]:
-        self.num_tokens += len(token_ids)
         for seq, token_id in zip(seqs, token_ids):
             seq.append_token(token_id)
             if (not seq.ignore_eos and token_id == self.eos) or seq.num_completion_tokens == seq.max_tokens:
                 seq.status = SequenceStatus.FINISHED
                 self.block_manager.deallocate(seq)
                 self.running.remove(seq)
-                self.num_finished += 1

nanovllm/engine/sequence.py

@@ -75,7 +75,7 @@ class Sequence:
         self.num_tokens += 1
 
     def __getstate__(self):
-        state = super().__getstate__()
+        state = vars(self).copy()
         if self.num_completion_tokens:
             state.pop("token_ids")
         return state

nanovllm/layers/embed_head.py

@@ -14,8 +14,8 @@ class VocabParallelEmbedding(nn.Module):
         embedding_dim: int,
     ):
         super().__init__()
-        self.tp_rank = 0  # get_tensor_model_parallel_rank()
+        self.tp_rank = dist.get_rank()
-        self.tp_size = 1  # get_tensor_model_parallel_world_size()
+        self.tp_size = dist.get_world_size()
         assert num_embeddings % self.tp_size == 0
         self.num_embeddings = num_embeddings
         self.num_embeddings_per_partition = self.num_embeddings // self.tp_size
@@ -39,7 +39,7 @@ class VocabParallelEmbedding(nn.Module):
             x = mask * (x - self.vocab_start_idx)
         y = F.embedding(x, self.weight)
         if self.tp_size > 1:
-            y = mask * y
+            y = mask.unsqueeze(1) * y
             dist.all_reduce(y)
         return y
 
@@ -65,8 +65,8 @@ class ParallelLMHead(VocabParallelEmbedding):
             last_indices = context.cu_seqlens_q[1:] - 1
             x = x[last_indices].contiguous()
         logits = F.linear(x, self.weight, self.bias)
-        # if self.tp_size > 1:
+        if self.tp_size > 1:
-        #     all_logits = [torch.empty_like(logits) for _ in range(self.tp_size)]
+            all_logits = [torch.empty_like(logits) for _ in range(self.tp_size)] if self.tp_rank == 0 else None
-        #     dist.gather(logits, all_logits, 0)
+            dist.gather(logits, all_logits, 0)
-        #     logits = torch.cat(all_logits, -1)
+            logits = torch.cat(all_logits, -1) if self.tp_rank == 0 else None
-        return logits if self.tp_rank == 0 else None
+        return logits

nanovllm/layers/linear.py

@@ -21,8 +21,8 @@ class LinearBase(nn.Module):
         self.input_size = input_size
         self.output_size = output_size
         self.tp_dim = tp_dim
-        self.tp_rank = 0  # get_tensor_model_parallel_rank()
+        self.tp_rank = dist.get_rank()
-        self.tp_size = 1  # get_tensor_model_parallel_world_size()
+        self.tp_size = dist.get_world_size()
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         raise NotImplementedError
@@ -65,7 +65,6 @@ class ColumnParallelLinear(LinearBase):
         self.input_size_per_partition = input_size
         self.output_size_per_partition = divide(output_size, self.tp_size)
         self.output_partition_sizes = [self.output_size_per_partition]
-        # If QKV or MergedColumn, use output size of each partition.
         if hasattr(self, "output_sizes"):
             self.output_partition_sizes = [
                 divide(output_size, self.tp_size)
@@ -101,8 +100,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         bias: bool = False,
     ):
         self.output_sizes = output_sizes
-        tp_size = 1  # get_tensor_model_parallel_world_size()
-        assert all(output_size % tp_size == 0 for output_size in output_sizes)
         super().__init__(input_size, sum(output_sizes), bias=bias)
 
     def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor, loaded_shard_id: int):
@@ -110,7 +107,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         shard_offset = sum(self.output_sizes[:loaded_shard_id]) // self.tp_size
         shard_size = self.output_sizes[loaded_shard_id] // self.tp_size
         param_data = param_data.narrow(self.tp_dim, shard_offset, shard_size)
-        # loaded_weight = loaded_weight.narrow(self.tp_dim, self.tp_rank * shard_size, shard_size)
+        loaded_weight = loaded_weight.chunk(self.tp_size, self.tp_dim)[self.tp_rank]
         assert param_data.size() == loaded_weight.size()
         param_data.copy_(loaded_weight)
 
@@ -131,8 +128,7 @@ class QKVParallelLinear(ColumnParallelLinear):
         if total_num_kv_heads is None:
             total_num_kv_heads = total_num_heads
         self.total_num_kv_heads = total_num_kv_heads
-        # Divide the weight matrix along the last dimension.
+        tp_size = dist.get_world_size()
-        tp_size = 1  # get_tensor_model_parallel_world_size()
         self.num_heads = divide(self.total_num_heads, tp_size)
         self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
         input_size = self.hidden_size
@@ -158,7 +154,7 @@ class QKVParallelLinear(ColumnParallelLinear):
             shard_size = self.num_kv_heads * self.head_size
             shard_offset = self.num_heads * self.head_size + self.num_kv_heads * self.head_size
         param_data = param_data.narrow(self.tp_dim, shard_offset, shard_size)
-        # loaded_weight = loaded_weight.narrow(self.tp_dim, self.tp_rank * shard_size, shard_size)
+        loaded_weight = loaded_weight.chunk(self.tp_size, self.tp_dim)[self.tp_rank]
         assert param_data.size() == loaded_weight.size()
         param_data.copy_(loaded_weight)
 

nanovllm/llm.py

@@ -2,4 +2,4 @@ from nanovllm.engine.llm_engine import LLMEngine
 
 
 class LLM(LLMEngine):
-    pass
+    pass

nanovllm/models/qwen3.py

@@ -1,5 +1,6 @@
 import torch
 from torch import nn
+import torch.distributed as dist
 from transformers import Qwen3Config
 
 from nanovllm.layers.activation import SiluAndMul
@@ -26,7 +27,7 @@ class Qwen3Attention(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
-        tp_size = 1  # get_tensor_model_parallel_world_size()
+        tp_size = dist.get_world_size()
         self.total_num_heads = num_heads
         assert self.total_num_heads % tp_size == 0
         self.num_heads = self.total_num_heads // tp_size