From bb57cb34e0bad70c6dea953320350066d9091d00 Mon Sep 17 00:00:00 2001 From: Jason Cheng Date: Wed, 21 Feb 2024 18:30:57 +0800 Subject: [PATCH] Added Qwen2 but generation is wrong removed unnecessary imports --- .../text_generation_server/models/__init__.py | 23 ++ .../custom_modeling/flash_qwen2_modeling.py | 390 ++++++++++++++++++ .../models/flash_qwen2.py | 77 ++++ 3 files changed, 490 insertions(+) create mode 100644 server/text_generation_server/models/custom_modeling/flash_qwen2_modeling.py create mode 100644 server/text_generation_server/models/flash_qwen2.py diff --git a/server/text_generation_server/models/__init__.py b/server/text_generation_server/models/__init__.py index da7d8416..65810b86 100644 --- a/server/text_generation_server/models/__init__.py +++ b/server/text_generation_server/models/__init__.py @@ -52,6 +52,9 @@ try: from text_generation_server.models.flash_llama import ( FlashLlama, ) + from text_generation_server.models.flash_qwen2 import ( + FlashQwen2, + ) from text_generation_server.models.flash_santacoder import ( FlashSantacoderSharded, ) @@ -75,6 +78,7 @@ if FLASH_ATTENTION: __all__.append(FlashMistral) __all__.append(FlashMixtral) __all__.append(FlashPhi) + __all__.append(FlashQwen2) MAMBA_AVAILABLE = True try: @@ -312,6 +316,25 @@ def get_model( dtype=dtype, trust_remote_code=trust_remote_code, ) + elif model_type == "qwen2": + if FLASH_ATTENTION: + return FlashQwen2( + model_id, + revision, + quantize=quantize, + dtype=dtype, + trust_remote_code=trust_remote_code, + ) + elif sharded: + raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Sharded Qwen2")) + else: + return CausalLM( + model_id, + revision, + quantize=quantize, + dtype=dtype, + trust_remote_code=trust_remote_code, + ) if model_type in ["RefinedWeb", "RefinedWebModel", "falcon"]: if sharded: diff --git a/server/text_generation_server/models/custom_modeling/flash_qwen2_modeling.py b/server/text_generation_server/models/custom_modeling/flash_qwen2_modeling.py new file mode 100644 index 00000000..849dd80e --- /dev/null +++ b/server/text_generation_server/models/custom_modeling/flash_qwen2_modeling.py @@ -0,0 +1,390 @@ +import torch +import torch.distributed + +from torch import nn +from transformers.activations import ACT2FN +from typing import Optional, List, Tuple + +from text_generation_server.utils import paged_attention, flash_attn +from text_generation_server.utils.layers import ( + TensorParallelRowLinear, + TensorParallelColumnLinear, + TensorParallelEmbedding, + PositionRotaryEmbedding, + TensorParallelHead, + get_linear, + FastRMSNorm, +) + + +def load_attention(config, prefix, weights): + if config.num_attention_heads != config.num_key_value_heads: + return _load_gqa(config, prefix, weights) + else: + return TensorParallelColumnLinear.load_multi( + config, + prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"], + dim=0, + weights=weights, + bias=False, + ) + + +def _load_gqa(config, prefix: str, weights): + assert config.hidden_size % config.num_attention_heads == 0 + assert config.num_attention_heads % weights.process_group.size() == 0 + + weight = weights.get_multi_weights_col( + prefixes=[f"{prefix}.q_proj", f"{prefix}.k_proj", f"{prefix}.v_proj"], + quantize=config.quantize, + dim=0, + ) + + if config.quantize not in ["gptq", "awq"]: + weight = weight.to(dtype=weights.dtype).to(device=weights.device) + + head_size = config.hidden_size // config.num_attention_heads + num_heads = config.num_attention_heads // weights.process_group.size() + num_key_value_heads = config.num_key_value_heads // weights.process_group.size() + assert list(weight.shape) == [ + (num_heads + 2 * num_key_value_heads) * head_size, + config.hidden_size, + ], f"{list(weight.shape)} != {[(num_heads + 2 * config.num_key_value_heads) * head_size, config.hidden_size]}" + + return TensorParallelColumnLinear( + get_linear(weight, bias=None, quantize=config.quantize) + ) + + +class Qwen2Attention(torch.nn.Module): + def __init__( + self, + prefix: str, + config, + weights, + ): + super().__init__() + self.max_past = ( + config.sliding_window if config.sliding_window is not None else -1 + ) + self.num_heads = config.num_attention_heads + self.hidden_size = config.hidden_size + self.head_size = self.hidden_size // self.num_heads + + self.rotary_emb = PositionRotaryEmbedding.static( + config=config, + dim=self.head_size, + base=config.rope_theta, + device=weights.device, + ) + + self.softmax_scale = self.head_size**-0.5 + + if self.num_heads % weights.process_group.size() != 0: + raise ValueError( + f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {self.num_heads} " + f"and `num_shards`: {weights.process_group.size()}" + ) + self.num_heads = self.num_heads // weights.process_group.size() + self.num_key_value_heads = ( + config.num_key_value_heads // weights.process_group.size() + ) + + self.query_key_value = load_attention(config, prefix, weights) + + self.o_proj = TensorParallelRowLinear.load( + config, + prefix=f"{prefix}.o_proj", + weights=weights, + bias=False, + ) + self.num_groups = self.num_heads // self.num_key_value_heads + self.kv_head_mapping = torch.arange( + 0, self.num_key_value_heads, dtype=torch.int32, device=weights.device + ).repeat_interleave(self.num_groups) + + def forward( + self, + hidden_states, + cos, + sin, + cu_seqlen_prefill, + kv_cache, + block_tables, + slots, + input_lengths, + max_s, + prefill_cache_indices, + ): + qkv = self.query_key_value(hidden_states) + query, kv = qkv.split( + [ + self.head_size * self.num_heads, + 2 * self.head_size * self.num_key_value_heads, + ], + dim=1, + ) + query = query.view(-1, self.num_heads, self.head_size) + kv = kv.view(-1, 2, self.num_key_value_heads, self.head_size) + + self.rotary_emb(query, torch.select(kv, dim=1, index=0), cos, sin) + + if prefill_cache_indices is not None: + kv_to_cache = kv[prefill_cache_indices] + else: + kv_to_cache = kv + + paged_attention.reshape_and_cache( + kv_to_cache[:, 0], kv_to_cache[:, 1], kv_cache[0], kv_cache[1], slots + ) + + # output tensor + attn_output = torch.empty_like(query) + + # Prefill + if cu_seqlen_prefill is not None: + # flash attention + flash_attn.attention( + query, + torch.select(kv, dim=1, index=0), + torch.select(kv, dim=1, index=1), + attn_output, + cu_seqlen_prefill, + max_s, + self.softmax_scale, + window_size_left=self.max_past, + ) + # Decode + else: + paged_attention.attention( + attn_output, + query, + kv_cache[0], + kv_cache[1], + self.kv_head_mapping, + self.softmax_scale, + block_tables, + input_lengths, + max_s, + ) + + return self.o_proj(attn_output.view(-1, self.num_heads * self.head_size)) + +class Qwen2MLP(nn.Module): + def __init__(self, prefix, config, weights): + super().__init__() + act = config.hidden_act + self.act = ( + ACT2FN[act] + if "gelu" not in act + else lambda x: torch.nn.functional.gelu( + x, + approximate=( + "tanh" if act in ["gelu_fast", "gelu_pytorch_tanh"] else "none" + ), + ) + ) + # Fuse gate and up proj + self.gate_up_proj = TensorParallelColumnLinear.load_multi( + config, + prefixes=[f"{prefix}.gate_proj", f"{prefix}.up_proj"], + weights=weights, + dim=0, + bias=False, + ) + self.down_proj = TensorParallelRowLinear.load( + config, + prefix=f"{prefix}.down_proj", + weights=weights, + bias=False, + ) + self.intermediate_size = ( + config.intermediate_size // weights.process_group.size() + ) + + def forward(self, hidden_states): + gate_up_states = self.gate_up_proj(hidden_states) + gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size) + return self.down_proj(self.act(gate_up_states[:, 0]) * gate_up_states[:, 1]) + + +class Qwen2Layer(nn.Module): + def __init__(self, layer_id, config, weights): + super().__init__() + prefix = f"model.layers.{layer_id}" + self.self_attn = Qwen2Attention(prefix=f"{prefix}.self_attn", config=config, weights=weights) + self.mlp = Qwen2MLP(prefix=f"{prefix}.mlp", config=config, weights=weights) + self.input_layernorm = FastRMSNorm.load( + prefix=f"{prefix}.input_layernorm", weights=weights, eps=config.rms_norm_eps + ) + self.post_attention_layernorm = FastRMSNorm.load( + prefix=f"{prefix}.post_attention_layernorm", + weights=weights, + eps=config.rms_norm_eps, + ) + + def forward( + self, + hidden_states, + residual, + cos, + sin, + cu_seqlen_prefill, + kv_cache, + block_tables, + slots, + input_lengths, + max_s, + prefill_cache_indices, + ): + normed_hidden_states, res = self.input_layernorm(hidden_states, residual) + + # Self Attention + attn_output = self.self_attn( + normed_hidden_states, + cos, + sin, + cu_seqlen_prefill, + kv_cache, + block_tables, + slots, + input_lengths, + max_s, + prefill_cache_indices, + ) + + # faster post attention rms norm + normed_attn_res_output, attn_res = self.post_attention_layernorm( + attn_output, res + ) + + mlp_output = self.mlp(normed_attn_res_output) + + return mlp_output, attn_res + +class Qwen2Model(torch.nn.Module): + def __init__(self, config, weights): + super().__init__() + process_group = weights.process_group + self.tp_rank = process_group.rank() + self.tp_world_size = process_group.size() + self.embed_tokens = TensorParallelEmbedding( + prefix="model.embed_tokens", weights=weights + ) + self.layers = nn.ModuleList( + [ + Qwen2Layer( + layer_id, + config, + weights, + ) + for layer_id in range(config.num_hidden_layers) + ] + ) + self.norm = FastRMSNorm.load( + prefix="model.norm", weights=weights, eps=config.rms_norm_eps + ) + + self.gradient_checkpointing = False + + self.head_size = self.layers[0].self_attn.head_size + self.num_heads = self.layers[0].self_attn.num_heads + self.num_key_value_heads = self.layers[0].self_attn.num_key_value_heads + + def forward( + self, + input_ids: torch.Tensor, + position_ids: torch.Tensor, + cu_seqlen_prefill: Optional[torch.Tensor], + kv_cache: List[Tuple[torch.Tensor, torch.Tensor]], + block_tables: torch.Tensor, + slots: torch.Tensor, + input_lengths: torch.Tensor, + max_s: int, + true_max_s: int, + prefill_cache_indices: Optional[torch.Tensor], + ) -> torch.Tensor: + hidden_states = self.embed_tokens(input_ids) + + # Get rotary cos and sin for this forward + # Avoid to index in each layer + cos, sin = self.layers[0].self_attn.rotary_emb.get_cos_sin( + position_ids, true_max_s, hidden_states.dtype + ) + + residual = None + for i, layer in enumerate(self.layers): + hidden_states, residual = layer( + hidden_states, + residual, + cos, + sin, + cu_seqlen_prefill, + kv_cache[i], + block_tables, + slots, + input_lengths, + max_s, + prefill_cache_indices, + ) + + hidden_states, _ = self.norm(hidden_states, residual) + + return hidden_states + + +class Qwen2ForCausalLM(torch.nn.Module): + def __init__(self, config, weights): + super().__init__() + + self.model = Qwen2Model(config, weights) + self.lm_head = TensorParallelHead.load( + config, + prefix="lm_head", + weights=weights, + ) + self.max_past = config.sliding_window + self.max_past_tensor = ( + torch.tensor(config.sliding_window, device=weights.device) + if self.max_past is not None + else None + ) + + def forward( + self, + input_ids: torch.Tensor, + position_ids: torch.Tensor, + cu_seqlen_prefill: Optional[torch.Tensor], + kv_cache: List[Tuple[torch.Tensor, torch.Tensor]], + block_tables: torch.Tensor, + slots: torch.Tensor, + input_lengths: torch.Tensor, + max_s: int, + prefill_cache_indices: Optional[torch.Tensor] = None, + lm_head_indices: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + true_max_s = max_s + if prefill_cache_indices is not None: + # Slots also need to be sliced as it has the same size as the whole kv tensor + slots = slots[prefill_cache_indices] + elif self.max_past is not None: + # Clamp in decode mode as paged attention requires clamped values whereas the flash attention + # kernel requires the true values + input_lengths = torch.clamp(input_lengths, max=self.max_past_tensor) + + hidden_states = self.model( + input_ids, + position_ids, + cu_seqlen_prefill, + kv_cache, + block_tables, + slots, + input_lengths, + max_s, + true_max_s, + prefill_cache_indices, + ) + if lm_head_indices is not None: + hidden_states = hidden_states[lm_head_indices] + logits = self.lm_head(hidden_states) + return logits diff --git a/server/text_generation_server/models/flash_qwen2.py b/server/text_generation_server/models/flash_qwen2.py new file mode 100644 index 00000000..791ebecd --- /dev/null +++ b/server/text_generation_server/models/flash_qwen2.py @@ -0,0 +1,77 @@ +import torch +import torch.distributed + +from opentelemetry import trace +from transformers import AutoTokenizer +from transformers.models.qwen2 import Qwen2Tokenizer +from typing import Optional + +from text_generation_server.models import FlashCausalLM +from text_generation_server.models.custom_modeling.flash_qwen2_modeling import ( + Qwen2ForCausalLM, +) +from transformers.models.qwen2 import Qwen2Config +from text_generation_server.utils import ( + initialize_torch_distributed, + weight_files, + Weights, +) + +tracer = trace.get_tracer(__name__) + + +class FlashQwen2(FlashCausalLM): + def __init__( + self, + model_id: str, + revision: Optional[str] = None, + quantize: Optional[str] = None, + dtype: Optional[torch.dtype] = None, + trust_remote_code: bool = False, + ): + self.process_group, rank, world_size = initialize_torch_distributed() + if torch.cuda.is_available(): + device = torch.device(f"cuda:{rank}") + dtype = torch.float16 if dtype is None else dtype + else: + raise NotImplementedError("FlashQwen2 is only available on GPU") + + try: + tokenizer = Qwen2Tokenizer.from_pretrained( + model_id, + revision=revision, + trust_remote_code=trust_remote_code, + ) + except Exception: + tokenizer = AutoTokenizer.from_pretrained( + model_id, + revision=revision, + trust_remote_code=trust_remote_code, + ) + + config = Qwen2Config.from_pretrained( + model_id, revision=revision, trust_remote_code=trust_remote_code + ) + config.quantize = quantize + + torch.distributed.barrier(group=self.process_group) + + filenames = weight_files(model_id, revision=revision, extension=".safetensors") + weights = Weights(filenames, device, dtype, process_group=self.process_group) + if config.quantize in ["gptq", "awq"]: + weights._set_gptq_params(model_id, revision) + + model = Qwen2ForCausalLM(config, weights) + + torch.distributed.barrier(group=self.process_group) + super(FlashQwen2, self).__init__( + model=model, + tokenizer=tokenizer, + num_layers=len(model.model.layers), + num_kv_heads=model.model.num_key_value_heads, + head_size=model.model.head_size, + dtype=dtype, + device=device, + rank=rank, + world_size=world_size, + )