Add support for GPT-NeoX (Pythia) (#50)

cacheflow/models/attention.py

@@ -150,20 +150,20 @@ class OPTCacheFlowAttention(GPTCacheFlowAttention):
         super().__init__(scale)
 
 
-class LlamaCacheFlowAttention(GPTCacheFlowAttention):
+class GPTNeoXCacheFlowAttention(GPTCacheFlowAttention):
-    """Llama uses GPT-NeoX style rotary embedding."""
+    """Attention with GPT-NeoX style rotary embedding."""
 
     def __init__(
         self,
         scale: float,
-        head_size: int,
+        rotary_dim: int,
         max_position: int = 8192,
         base: int = 10000,
     ) -> None:
         super().__init__(scale)
 
         # Create the cos and sin cache.
-        inv_freq = 1.0 / (base ** (torch.arange(0, head_size, 2) / head_size))
+        inv_freq = 1.0 / (base ** (torch.arange(0, rotary_dim, 2) / rotary_dim))
         t = torch.arange(max_position).float()
         freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
         cos = freqs.cos()
@@ -174,7 +174,7 @@ class LlamaCacheFlowAttention(GPTCacheFlowAttention):
         # initializing the model. Make it more robust.
         torch_dtype = torch.get_default_dtype()
         cache = cache.to(torch_dtype)
-        # Embedding size: [max_position, head_size]
+        # Embedding size: [max_position, rotary_dim]
         self.register_buffer('cos_sin_cache', cache, persistent=False)
 
     def forward(
@@ -190,10 +190,12 @@ class LlamaCacheFlowAttention(GPTCacheFlowAttention):
     ) -> torch.Tensor:                          # [num_tokens, num_heads * head_size]
         # Apply rotary embedding to the query and key before passing them
         # to the attention op.
+        head_size = value_cache.shape[2]
         pos_encoding_ops.rotary_embedding_neox(
             positions,
             query,
             key,
+            head_size,
             self.cos_sin_cache,
         )
         return super().forward(
@@ -205,3 +207,7 @@ class LlamaCacheFlowAttention(GPTCacheFlowAttention):
             input_metadata,
             cache_event,
         )
+
+
+class LlamaCacheFlowAttention(GPTNeoXCacheFlowAttention):
+    """LLaMA uses the GPT-NeoX style rotary embedding."""

cacheflow/models/gpt_neox.py

@@ -0,0 +1,278 @@
+"""1D GPT-NeoX model compatible with HuggingFace weights."""
+import os
+import glob
+import filelock
+from tqdm import tqdm
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from torch import nn
+from huggingface_hub import snapshot_download
+
+from cacheflow.models import InputMetadata
+from cacheflow.models.attention import GPTNeoXCacheFlowAttention
+from cacheflow.models.sample import Sampler
+from cacheflow.parallel_utils.parallel_state import (
+    get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
+from cacheflow.parallel_utils.tensor_parallel import (VocabParallelEmbedding,
+                                                      ColumnParallelLinear,
+                                                      RowParallelLinear)
+from cacheflow.sequence import SequenceOutputs
+
+KVCache = Tuple[torch.Tensor, torch.Tensor]
+
+
+class GPTNeoXAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+
+        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = self.total_num_heads // tensor_model_parallel_world_size
+
+        self.query_key_value = ColumnParallelLinear(config.hidden_size,
+                                                    3 * config.hidden_size,
+                                                    gather_output=False,
+                                                    perform_initialization=False)
+        self.dense = RowParallelLinear(config.hidden_size, config.hidden_size,
+                                       input_is_parallel=True,
+                                       perform_initialization=False)
+
+        scaling = self.head_size ** -0.5
+        rotary_dim = int(self.head_size * config.rotary_pct)
+        assert rotary_dim % 2 == 0
+        self.attn = GPTNeoXCacheFlowAttention(scaling, rotary_dim)
+
+    def forward(
+        self,
+        position_ids: torch.LongTensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+        cache_event: Optional[torch.cuda.Event],
+    ) -> torch.Tensor:
+        qkv, _ = self.query_key_value(hidden_states)
+
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        k_cache, v_cache = kv_cache
+        attn_output = self.attn(
+            position_ids, q, k, v, k_cache, v_cache, input_metadata, cache_event)
+        output, _ = self.dense(attn_output)
+        return output
+
+
+class GPTNeoXMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense_h_to_4h = ColumnParallelLinear(config.hidden_size,
+                                                  config.intermediate_size,
+                                                  gather_output=False,
+                                                  perform_initialization=False)
+        self.dense_4h_to_h = RowParallelLinear(config.intermediate_size, config.hidden_size,
+                                               input_is_parallel=True,
+                                               perform_initialization=False)
+        if config.hidden_act != 'gelu':
+            raise ValueError(f'Unsupported activation: {config.hidden_act}. '
+                             'Only gelu is supported for now.')
+        self.act = torch.nn.GELU()
+
+    def forward(self, hidden_states):
+        hidden_states, _ = self.dense_h_to_4h(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.dense_4h_to_h(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXLayer(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.use_parallel_residual = config.use_parallel_residual
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = GPTNeoXAttention(config)
+        self.mlp = GPTNeoXMLP(config)
+
+    def forward(
+        self,
+        position_ids: torch.LongTensor,
+        hidden_states: torch.Tensor,
+        kv_cache: KVCache,
+        input_metadata: InputMetadata,
+        cache_event: Optional[torch.cuda.Event],
+    ) -> torch.Tensor:
+        attn_input = self.input_layernorm(hidden_states)
+        attn_output = self.attention(
+            position_ids=position_ids,
+            hidden_states=attn_input,
+            kv_cache=kv_cache,
+            input_metadata=input_metadata,
+            cache_event=cache_event,
+        )
+
+        if self.use_parallel_residual:
+            # pseudocode:
+            # x = x + attn(ln1(x)) + mlp(ln2(x))
+            mlp_input = self.post_attention_layernorm(hidden_states)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output + hidden_states
+        else:
+            # pseudocode:
+            # x = x + attn(ln1(x))
+            # x = x + mlp(ln2(x))
+            attn_output = attn_output + hidden_states
+            mlp_input = self.post_attention_layernorm(attn_output)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output
+        return hidden_states
+
+
+class GPTNeoXModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.embed_in = VocabParallelEmbedding(config.vocab_size, config.hidden_size,
+                                               perform_initialization=False)
+        self.layers = nn.ModuleList([GPTNeoXLayer(config) for _ in range(config.num_hidden_layers)])
+        self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        position_ids: torch.LongTensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+        cache_events: Optional[List[torch.cuda.Event]],
+    ) -> torch.Tensor:
+        hidden_states = self.embed_in(input_ids)
+        for i in range(len(self.layers)):
+            if cache_events is None:
+                cache_event = None
+            else:
+                cache_event = cache_events[i]
+            layer = self.layers[i]
+            hidden_states = layer(
+                position_ids,
+                hidden_states,
+                kv_caches[i],
+                input_metadata,
+                cache_event,
+            )
+        hidden_states = self.final_layer_norm(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXForCausalLM(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.gpt_neox = GPTNeoXModel(config)
+        self.embed_out = ColumnParallelLinear(config.hidden_size, config.vocab_size,
+                                              bias=False, gather_output=False,
+                                              perform_initialization=False)
+        self.sampler = Sampler()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        positions: torch.LongTensor,
+        kv_caches: List[KVCache],
+        input_metadata: InputMetadata,
+        cache_events: Optional[List[torch.cuda.Event]],
+    ) -> Dict[int, SequenceOutputs]:
+        hidden_states = self.gpt_neox(
+            input_ids, positions, kv_caches, input_metadata, cache_events)
+        next_tokens = self.sampler(
+            self.embed_out.weight, hidden_states, input_metadata)
+        return next_tokens
+
+    _column_parallel_weights = ["embed_in.weight", "embed_out.weight", "dense_h_to_4h.weight", "dense_h_to_4h.bias"]
+    _row_parallel_weights = ["dense.weight", "dense_4h_to_h.weight"]
+
+    def load_weights(self, weights_path: str):
+        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
+        state_dict = self.state_dict()
+        for name, param in state_dict.items():
+            if "query_key_value" in name:
+                # NOTE(woosuk): GPT-NeoX's fused QKV has the shape of
+                # [num_heads * 3 * head_size, num_heads * head_size], while the
+                # required shape is [3 * num_heads * head_size, num_heads * head_size].
+                # Thus, we need weight conversion.
+                loaded_weight = torch.from_numpy(
+                    np.load(os.path.join(weights_path, name)))
+                shard_size = param.shape[0]
+                loaded_weight = loaded_weight[shard_size * tensor_model_parallel_rank
+                                              :shard_size * (tensor_model_parallel_rank + 1)]
+
+                num_heads = self.config.num_attention_heads
+                hidden_size = self.config.hidden_size
+                head_size = hidden_size // num_heads
+                if 'query_key_value.weight' in name:
+                    loaded_weight = loaded_weight.view(-1, 3, head_size, hidden_size)
+                    loaded_weight = loaded_weight.transpose(0, 1)
+                    loaded_weight = loaded_weight.reshape(-1, hidden_size).contiguous()
+                elif 'query_key_value.bias' in name:
+                    loaded_weight = loaded_weight.view(-1, 3, head_size)
+                    loaded_weight = loaded_weight.transpose(0, 1)
+                    loaded_weight = loaded_weight.reshape(-1).contiguous()
+                else:
+                    assert False
+            else:
+                loaded_weight = torch.from_numpy(
+                    np.load(os.path.join(weights_path, name)))
+                for p in self._column_parallel_weights:
+                    if p in name:
+                        shard_size = param.shape[0]
+                        loaded_weight = loaded_weight[
+                            shard_size * tensor_model_parallel_rank
+                            :shard_size * (tensor_model_parallel_rank + 1)]
+                        break
+                for p in self._row_parallel_weights:
+                    if p in name:
+                        shard_size = param.shape[1]
+                        loaded_weight = loaded_weight[
+                            :,
+                            shard_size * tensor_model_parallel_rank
+                            :shard_size * (tensor_model_parallel_rank + 1)]
+                        break
+
+            assert param.shape == loaded_weight.shape
+            param.data.copy_(loaded_weight)
+
+    @staticmethod
+    def get_weights(model_name: str, path: str):
+        path = os.path.join(path, f"{model_name}-np")
+        path = os.path.abspath(os.path.expanduser(path))
+        os.makedirs(path, exist_ok=True)
+        lock_path = os.path.join(path, "file_lock")
+        lock = filelock.FileLock(lock_path)
+
+        with lock:
+            test_weight_path = os.path.join(
+                path, "gpt_neox.embed_in.weight")
+            if os.path.exists(test_weight_path):
+                return path
+
+            folder = snapshot_download(model_name, allow_patterns="*.bin",
+                                       cache_dir=os.path.join(path, "cache"))
+            bin_files = glob.glob(os.path.join(folder, "*.bin"))
+
+            for bin_file in tqdm(bin_files, desc="Convert format"):
+                state = torch.load(bin_file, map_location="cpu")
+                for name, param in tqdm(state.items(), leave=False):
+                    param_path = os.path.join(path, name)
+                    with open(param_path, "wb") as f:
+                        np.save(f, param.cpu().detach().numpy())
+
+            return path
+
+    def initialize_dummy_weights(self) -> None:
+        for param in self.state_dict().values():
+            param.data.uniform_(-1e-3, 1e-3)

cacheflow/models/llama.py

@@ -289,4 +289,4 @@ class LlamaForCausalLM(nn.Module):
 
     def initialize_dummy_weights(self) -> None:
         for param in self.state_dict().values():
-            param.data.uniform_(-0.1, 0.1)
+            param.data.uniform_(-1e-3, 1e-3)

cacheflow/models/memory_analyzer.py

@@ -40,6 +40,37 @@ class CacheFlowMemoryAnalyzer:
         max_num_blocks = swap_space // self.get_cache_block_size()
         return max_num_blocks
 
+    def get_param_size(self) -> int:
+        raise NotImplementedError()
+
+    def get_max_act_size(self, max_num_batched_tokens: int) -> int:
+        raise NotImplementedError()
+
+    def get_cache_block_size(self) -> int:
+        key_cache_block = self.block_size * self.hidden_size // self.tensor_parallel_size
+        value_cache_block = key_cache_block
+        total = self.num_layers * (key_cache_block + value_cache_block)
+        dtype_size = get_dtype_size(self.dtype)
+        return dtype_size * total
+
+    def get_max_num_gpu_blocks(
+        self,
+        max_num_batched_tokens: int,
+        memory_utilization: float = 0.95,
+    ) -> int:
+        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
+        usable_memory = int(memory_utilization * self.gpu_memory)
+
+        param_size = self.get_param_size()
+        act_size = self.get_max_act_size(max_num_batched_tokens)
+        workspace_size = self.get_workspace_size()
+
+        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
+        if max_cache_size <= 0:
+            raise RuntimeError('Not enough GPU memory.')
+        max_num_blocks = max_cache_size // self.get_cache_block_size()
+        return max_num_blocks
+
 
 class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
 
@@ -69,7 +100,7 @@ class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         self.vocab_size = config.vocab_size
         self.max_position = config.max_position_embeddings
 
-    def _get_param_size(self) -> int:
+    def get_param_size(self) -> int:
         word_embedding = self.vocab_size * self.embedding_size // self.tensor_parallel_size
         if self.embedding_size != self.hidden_size:
             # Project in/out.
@@ -93,7 +124,7 @@ class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         dtype_size = get_dtype_size(self.dtype)
         return dtype_size * total
 
-    def _get_max_act_size(
+    def get_max_act_size(
         self,
         max_num_batched_tokens: int,
     ) -> int:
@@ -114,31 +145,6 @@ class OPTMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         dtype_size = get_dtype_size(self.dtype)
         return dtype_size * max_act
 
-    def get_cache_block_size(self) -> int:
-        key_cache_block = self.block_size * self.hidden_size // self.tensor_parallel_size
-        value_cache_block = key_cache_block
-        total = self.num_layers * (key_cache_block + value_cache_block)
-        dtype_size = get_dtype_size(self.dtype)
-        return dtype_size * total
-
-    def get_max_num_gpu_blocks(
-        self,
-        max_num_batched_tokens: int,
-        memory_utilization: float = 0.95,
-    ) -> int:
-        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
-        usable_memory = int(memory_utilization * self.gpu_memory)
-
-        param_size = self._get_param_size()
-        act_size = self._get_max_act_size(max_num_batched_tokens)
-        workspace_size = self.get_workspace_size()
-
-        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
-        if max_cache_size <= 0:
-            raise RuntimeError('Not enough GPU memory.')
-        max_num_blocks = max_cache_size // self.get_cache_block_size()
-        return max_num_blocks
-
 
 class LlamaMemoryAnalyzer(CacheFlowMemoryAnalyzer):
 
@@ -167,9 +173,10 @@ class LlamaMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         self.vocab_size = config.vocab_size
         self.max_position = 8192
 
-    def _get_param_size(self) -> int:
+    def get_param_size(self) -> int:
+        # NOTE: LLaMA does not tie the two embeddings.
         word_embedding = self.vocab_size * self.hidden_size // self.tensor_parallel_size
-        position_embedding = self.max_position * self.hidden_size
+        lm_head = self.vocab_size * self.hidden_size // self.tensor_parallel_size
 
         # NOTE: LLaMA does not have bias terms.
         ln1 = self.hidden_size
@@ -188,11 +195,11 @@ class LlamaMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         up = self.hidden_size * self.ffn_size // self.tensor_parallel_size
         ffn = ln2 + gate + down + up
 
-        total = (word_embedding + position_embedding + self.num_layers * (mha + ffn))
+        total = word_embedding + self.num_layers * (mha + ffn) + lm_head
         dtype_size = get_dtype_size(self.dtype)
         return dtype_size * total
 
-    def _get_max_act_size(
+    def get_max_act_size(
         self,
         max_num_batched_tokens: int,
     ) -> int:
@@ -213,28 +220,78 @@ class LlamaMemoryAnalyzer(CacheFlowMemoryAnalyzer):
         dtype_size = get_dtype_size(self.dtype)
         return dtype_size * max_act
 
-    def get_cache_block_size(self) -> int:
+
-        key_cache_block = self.block_size * self.hidden_size // self.tensor_parallel_size
+class GPTNeoXMemoryAnalyzer(CacheFlowMemoryAnalyzer):
-        value_cache_block = key_cache_block
+
-        total = self.num_layers * (key_cache_block + value_cache_block)
+    def __init__(
+        self,
+        model_name: str,
+        block_size: int,
+        dtype: torch.dtype,
+        gpu_memory: int,
+        cpu_memory: int,
+        tensor_parallel_size: int,
+    ) -> None:
+        self.model_name = model_name
+        self.block_size = block_size
+        self.dtype = dtype
+        self.gpu_memory = gpu_memory
+        self.cpu_memory = cpu_memory
+        self.tensor_parallel_size = tensor_parallel_size
+
+        config = AutoConfig.from_pretrained(model_name)
+        self.num_layers = config.num_hidden_layers
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_size = config.hidden_size // self.num_heads
+        self.ffn_size = config.intermediate_size
+        self.vocab_size = config.vocab_size
+        self.max_position = 8192
+        self.tie_word_embeddings = config.tie_word_embeddings
+
+    def get_param_size(self) -> int:
+        word_embedding = self.vocab_size * self.hidden_size // self.tensor_parallel_size
+        if self.tie_word_embeddings:
+            lm_head = 0
+        else:
+            lm_head = self.vocab_size * self.hidden_size // self.tensor_parallel_size
+
+        ln1 = 2 * self.hidden_size
+        q = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
+        k = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
+        v = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
+        out = self.hidden_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
+        # Rotary embedding.
+        # TODO(woosuk): Share the rotary embedding between layers.
+        rot = self.max_position * self.head_size
+        mha = ln1 + q + k + v + out + rot
+
+        ln2 = 2 * self.hidden_size
+        ffn1 = self.hidden_size * self.ffn_size // self.tensor_parallel_size + self.ffn_size
+        ffn2 = self.ffn_size * self.hidden_size // self.tensor_parallel_size + self.hidden_size
+        ffn = ln2 + ffn1 + ffn2
+
+        total = word_embedding + self.num_layers * (mha + ffn) + lm_head
         dtype_size = get_dtype_size(self.dtype)
         return dtype_size * total
 
-    def get_max_num_gpu_blocks(
+    def get_max_act_size(
         self,
         max_num_batched_tokens: int,
-        memory_utilization: float = 0.95,
     ) -> int:
-        # NOTE(woosuk): This assumes that the machine has homogeneous GPUs.
+        # NOTE: We approxmiately calculate the maximum activation size by
-        gpu_memory = self.gpu_memory
+        # estimating
-        usable_memory = int(memory_utilization * gpu_memory)
+        # 1) the maximum activation tensor size during inference
-
+        # 2) the residual tensor size during inference
-        param_size = self._get_param_size()
+        # Here, we assume that FlashAttention is used and
-        act_size = self._get_max_act_size(max_num_batched_tokens)
+        # thus the attention maps are never materialized in GPU DRAM.
-        workspace_size = self.get_workspace_size()
+        residual = max_num_batched_tokens * self.hidden_size
-
+        qkv = 3 * (max_num_batched_tokens * self.hidden_size) // self.tensor_parallel_size
-        max_cache_size = usable_memory - (param_size + act_size + workspace_size)
+        ffn = 2 * (max_num_batched_tokens * self.ffn_size) // self.tensor_parallel_size
-        if max_cache_size <= 0:
+        # Double the activation size for input and output.
-            raise RuntimeError('Not enough GPU memory.')
+        max_act = 2 * (max(qkv, ffn) + residual)
-        max_num_blocks = max_cache_size // self.get_cache_block_size()
+        # Size of output logits.
-        return max_num_blocks
+        output_logits = 2 * (max_num_batched_tokens * self.vocab_size)
+        max_act = max(max_act, output_logits)
+        dtype_size = get_dtype_size(self.dtype)
+        return dtype_size * max_act

cacheflow/models/model_utils.py

@@ -1,13 +1,14 @@
 from typing import Union
 
-import numpy as np
 import torch
 import torch.nn as nn
 from transformers import AutoConfig
 
 from cacheflow.models.memory_analyzer import CacheFlowMemoryAnalyzer
+from cacheflow.models.memory_analyzer import GPTNeoXMemoryAnalyzer
 from cacheflow.models.memory_analyzer import LlamaMemoryAnalyzer
 from cacheflow.models.memory_analyzer import OPTMemoryAnalyzer
+from cacheflow.models.gpt_neox import GPTNeoXForCausalLM
 from cacheflow.models.llama import LlamaForCausalLM
 from cacheflow.models.opt import OPTForCausalLM
 from cacheflow.models.utils import get_torch_dtype
@@ -16,11 +17,15 @@ from cacheflow.models.utils import get_torch_dtype
 _MODELS = {
     'llama': LlamaForCausalLM,
     'opt': OPTForCausalLM,
+    'stablelm': GPTNeoXForCausalLM,
+    'pythia': GPTNeoXForCausalLM,
 }
 
 _MEMORY_ANALYZERS = {
     'llama': LlamaMemoryAnalyzer,
     'opt': OPTMemoryAnalyzer,
+    'stablelm': GPTNeoXMemoryAnalyzer,
+    'pythia': GPTNeoXMemoryAnalyzer,
 }
 
 

cacheflow/models/opt.py

@@ -327,4 +327,4 @@ class OPTForCausalLM(nn.Module):
 
     def initialize_dummy_weights(self) -> None:
         for param in self.state_dict().values():
-            param.data.uniform_(-0.1, 0.1)
+            param.data.uniform_(-1e-3, 1e-3)

csrc/pos_encoding.cpp

@@ -4,6 +4,7 @@ void rotary_embedding_neox(
   torch::Tensor& positions,
   torch::Tensor& query,
   torch::Tensor& key,
+  int head_size,
   torch::Tensor& cos_sin_cache);
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {

csrc/pos_encoding_kernels.cu

@@ -8,16 +8,17 @@ __global__ void rotary_embedding_neox_kernel(
   const int64_t* __restrict__ positions,        // [num_tokens]
   scalar_t* __restrict__ query,                 // [num_tokens, num_heads, head_size]
   scalar_t* __restrict__ key,                   // [num_tokens, num_heads, head_size]
-  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, head_size // 2]
+  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, rot_dim // 2]
+  const int rot_dim,
   const int stride,
   const int num_heads,
   const int head_size) {
   // Each thread block is responsible for one token.
   const int token_idx = blockIdx.x;
   int64_t pos = positions[token_idx];
-  const scalar_t* cache_ptr = cos_sin_cache + pos * head_size;
+  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
 
-  const int embed_dim = head_size / 2;
+  const int embed_dim = rot_dim / 2;
   const int n = num_heads * embed_dim;
   for (int i = threadIdx.x; i < n; i += blockDim.x) {
     const int head_idx = i / embed_dim;
@@ -51,16 +52,17 @@ void rotary_embedding_neox(
   torch::Tensor& positions,         // [num_tokens]
   torch::Tensor& query,             // [num_tokens, num_heads * head_size]
   torch::Tensor& key,               // [num_tokens, num_heads * head_size]
-  torch::Tensor& cos_sin_cache)     // [max_position, head_size]
+  int head_size,
+  torch::Tensor& cos_sin_cache)     // [max_position, rot_dim]
 {
   int num_tokens = query.size(0);
-  int head_size = cos_sin_cache.size(1);
+  int rot_dim = cos_sin_cache.size(1);
   int num_heads = query.size(1) / head_size;
   int stride = query.stride(0);
   TORCH_CHECK(stride == key.stride(0));
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(num_heads * head_size / 2, 512));
+  dim3 block(std::min(num_heads * rot_dim / 2, 512));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
     query.scalar_type(),
@@ -71,6 +73,7 @@ void rotary_embedding_neox(
         query.data_ptr<scalar_t>(),
         key.data_ptr<scalar_t>(),
         cos_sin_cache.data_ptr<scalar_t>(),
+        rot_dim,
         stride,
         num_heads,
         head_size);

tests/kernels/pos_encoding.py

@@ -34,6 +34,7 @@ class RefRotaryEmbeddingNeox(nn.Module):
         base: int = 10000,
     ) -> None:
         super().__init__()
+        self.rotary_dim = dim
         self.max_position_embeddings = max_position_embeddings
 
         # Create cos and sin embeddings.
@@ -52,13 +53,24 @@ class RefRotaryEmbeddingNeox(nn.Module):
         query: torch.Tensor,            # [num_tokens, num_heads, head_size]
         key: torch.Tensor,              # [num_tokens, num_heads, head_size]
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+
+        query_rot = query[..., : self.rotary_dim]
+        query_pass = query[..., self.rotary_dim :]
+        key_rot = key[..., : self.rotary_dim]
+        key_pass = key[..., self.rotary_dim :]
+
+
+        query_rot = query_rot.transpose(0, 1)
+        key_rot = key_rot.transpose(0, 1)
         cos = F.embedding(positions, self.cos_cached)
         sin = F.embedding(positions, self.sin_cached)
-        query = query.transpose(0, 1)
+        query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
-        key = key.transpose(0, 1)
+        query_rot = query_rot.transpose(0, 1).contiguous()
-        query, key = apply_rotary_pos_emb(query, key, cos, sin)
+        key_rot = key_rot.transpose(0, 1).contiguous()
-        query = query.transpose(0, 1).contiguous()
+
-        key = key.transpose(0, 1).contiguous()
+        query = torch.cat((query_rot, query_pass), dim=-1)
+        key = torch.cat((key_rot, key_pass), dim=-1)
+
         # Output query/key shape: [num_tokens, num_tokens, head_size]
         return query, key
 
@@ -69,6 +81,7 @@ def test_rotary_embedding_neox(
     num_heads: int,
     head_size: int,
     max_position: int,
+    rotary_dim: int,
     dtype: torch.dtype,
     base: int = 10000,
 ) -> None:
@@ -77,7 +90,7 @@ def test_rotary_embedding_neox(
     key = torch.randn(num_tokens, num_heads * head_size, dtype=dtype, device='cuda')
 
     # Create the rotary embedding.
-    inv_freq = 1.0 / (base ** (torch.arange(0, head_size, 2) / head_size))
+    inv_freq = 1.0 / (base ** (torch.arange(0, rotary_dim, 2) / rotary_dim))
     t = torch.arange(max_position).float()
     freqs = torch.einsum('i,j -> ij', t, inv_freq.float())
     cos = freqs.cos()
@@ -92,12 +105,13 @@ def test_rotary_embedding_neox(
         positions,
         out_query,
         out_key,
+        head_size,
         cos_sin_cache,
     )
 
     # Run the reference implementation.
     ref_rotary_embedding = RefRotaryEmbeddingNeox(
-        dim=head_size,
+        dim=rotary_dim,
         max_position_embeddings=max_position,
         base=base,
     ).to(dtype=dtype, device='cuda')
@@ -123,5 +137,6 @@ if __name__ == '__main__':
                 num_heads=5,
                 head_size=head_size,
                 max_position=8192,
+                rotary_dim=int(head_size * 0.25),
                 dtype=dtype,
             )