sglang 0.2.14.post1__py3-none-any.whl → 0.2.15__py3-none-any.whl

sglang/srt/models/deepseek_v2.py

@@ -19,6 +19,7 @@ limitations under the License.
 from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
+from flashinfer import bmm_fp8
 from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
@@ -45,6 +46,7 @@ from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
@@ -160,6 +162,15 @@ def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
     return 0.1 * mscale * math.log(scale) + 1.0
 
 
+def input_to_float8(x, dtype=torch.float8_e4m3fn):
+    finfo = torch.finfo(dtype)
+    min_val, max_val = x.aminmax()
+    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    scale = finfo.max / amax
+    x_scl_sat = (x * scale).clamp(min=finfo.min, max=finfo.max)
+    return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
+
+
 class DeepseekV2Attention(nn.Module):
 
     def __init__(
@@ -254,11 +265,6 @@ class DeepseekV2Attention(nn.Module):
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        # self.attn = Attention(self.num_heads,
-        #                       self.qk_head_dim,
-        #                       self.scaling,
-        #                       num_kv_heads=self.num_heads)
-
         # TODO, support head_size 192
         self.attn = RadixAttention(
             self.num_local_heads,
@@ -282,7 +288,7 @@ class DeepseekV2Attention(nn.Module):
             q = self.q_proj(hidden_states)[0].view(
                 -1, self.num_local_heads, self.qk_head_dim
             )
-        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
         latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
         latent_cache = latent_cache.unsqueeze(1)
@@ -416,12 +422,9 @@ class DeepseekV2AttentionMLA(nn.Module):
             v_head_dim=self.kv_lora_rank,
         )
 
-        kv_b_proj = self.kv_b_proj
-        w_kc, w_vc = kv_b_proj.weight.unflatten(
-            0, (-1, qk_nope_head_dim + v_head_dim)
-        ).split([qk_nope_head_dim, v_head_dim], dim=1)
-        self.w_kc = w_kc
-        self.w_vc = w_vc
+        self.w_kc = None
+        self.w_vc = None
+        self.w_scale = None
 
     def forward(
         self,
@@ -442,8 +445,17 @@ class DeepseekV2AttentionMLA(nn.Module):
                 -1, self.num_local_heads, self.qk_head_dim
             )
         q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
-        q_nope_out = q_input[..., : self.kv_lora_rank]
-        torch.bmm(q_nope.transpose(0, 1), self.w_kc, out=q_nope_out.transpose(0, 1))
+
+        if self.w_kc.dtype == torch.float8_e4m3fn:
+            q_nope_val, q_nope_scale = input_to_float8(
+                q_nope.transpose(0, 1), torch.float8_e4m3fn
+            )
+            q_nope_out = bmm_fp8(
+                q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
+            )
+        else:
+            q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
+        q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
 
         latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
         v_input = latent_cache[..., : self.kv_lora_rank]
@@ -458,16 +470,21 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         attn_output = self.attn(q_input, k_input, v_input, input_metadata)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
-        attn_bmm_output = attn_output.new_empty(
-            q_len, self.num_local_heads, self.v_head_dim
-        )
-        torch.bmm(
-            attn_output.transpose(0, 1),
-            self.w_vc.transpose(1, 2).contiguous(),
-            out=attn_bmm_output.transpose(0, 1),
-        )
 
-        attn_output = attn_bmm_output.flatten(1, 2)
+        if self.w_vc.dtype == torch.float8_e4m3fn:
+            attn_output_val, attn_output_scale = input_to_float8(
+                attn_output.transpose(0, 1), torch.float8_e4m3fn
+            )
+            attn_bmm_output = bmm_fp8(
+                attn_output_val,
+                self.w_vc,
+                attn_output_scale,
+                self.w_scale,
+                torch.bfloat16,
+            )
+        else:
+            attn_bmm_output = torch.bmm(attn_output.transpose(0, 1), self.w_vc)
+        attn_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
         output, _ = self.o_proj(attn_output)
 
         return output
@@ -632,6 +649,7 @@ class DeepseekV2ForCausalLM(nn.Module):
             config.vocab_size, config.hidden_size, quant_config=quant_config
         )
         self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
 
     def forward(
         self,
@@ -640,9 +658,11 @@ class DeepseekV2ForCausalLM(nn.Module):
         input_metadata: InputMetadata,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata)
-        return self.logits_processor(
+        logits_output = self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return sample_output, logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
@@ -695,7 +715,7 @@ class DeepseekV2ForCausalLM(nn.Module):
                     weight_loader(
                         param,
                         loaded_weight,
-                        weight_name,
+                        name,
                         shard_id=shard_id,
                         expert_id=expert_id,
                     )
@@ -711,5 +731,17 @@ class DeepseekV2ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
+        if global_server_args_dict["enable_mla"]:
+            for layer_id in range(self.config.num_hidden_layers):
+                self_attn = self.model.layers[layer_id].self_attn
+                w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(
+                    0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
+                ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
+                self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
+                self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
+                if hasattr(self_attn.kv_b_proj, "weight_scale"):
+                    self_attn.w_scale = self_attn.kv_b_proj.weight_scale
+                del self_attn.kv_b_proj
+
 
 EntryClass = DeepseekV2ForCausalLM

sglang/srt/models/exaone.py

@@ -0,0 +1,399 @@
+"""
+Copyright 2024 The LGcns AI Engineering Team
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+# Adapted from llama2.py
+"""Inference-only Exaone model compatible with HuggingFace weights."""
+
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from vllm.config import CacheConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
+from sglang.srt.model_executor.forward_batch_info import InputMetadata
+
+
+class ExaoneGatedMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.c_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.c_proj(x)
+        return x
+
+
+class ExaoneAttention(nn.Module):
+    def __init__(
+        self,
+        config,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 500000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_is_neox_style: bool = True,
+        max_position_embeddings: int = 4096,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_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
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(
+            config, "head_dim", self.hidden_size // self.total_num_heads
+        )
+        self.rotary_dim = int(
+            self.head_dim * getattr(config, "partial_rotary_factor", 1)
+        )
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.out_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_neox_style=rope_is_neox_style,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class ExaoneDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 500000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+            config, "original_max_position_embeddings", None
+        ):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings
+            )
+        rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 4096)
+        self.self_attn = ExaoneAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            rope_is_neox_style=rope_is_neox_style,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = ExaoneGatedMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.activation_function,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+        )
+        rms_norm_eps = config.layer_norm_epsilon
+        self.ln_1 = RMSNorm(config.hidden_size, eps=rms_norm_eps)
+        self.ln_2 = RMSNorm(config.hidden_size, eps=rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.ln_1(hidden_states)
+        else:
+            hidden_states, residual = self.ln_1(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ln_2(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class ExaoneModel(nn.Module):
+    def __init__(
+        self,
+        config,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.h = nn.ModuleList(
+            [
+                ExaoneDecoderLayer(
+                    config, i, quant_config=quant_config, prefix=f"model.h.{i}"
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        rms_norm_eps = config.layer_norm_epsilon
+        self.ln_f = RMSNorm(config.hidden_size, eps=rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.wte(input_ids)
+        else:
+            hidden_states = input_embeds
+        residual = None
+        for i in range(len(self.h)):
+            layer = self.h[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.ln_f(hidden_states, residual)
+        return hidden_states
+
+
+class ExaoneForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+        efficient_weight_load=False,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.transformer = ExaoneModel(config, quant_config=quant_config)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> LogitsProcessorOutput:
+        hidden_states = self.transformer(
+            input_ids, positions, input_metadata, input_embeds
+        )
+        logits_output = self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return sample_output, logits_output
+
+    def get_module_name(self, name):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id, num_shard)
+            ("qkv_proj", "q_proj", "q", 3),
+            ("qkv_proj", "k_proj", "k", 3),
+            ("qkv_proj", "v_proj", "v", 3),
+            ("gate_up_proj", "c_fc_0", 0, 2),
+            ("gate_up_proj", "c_fc_1", 1, 2),
+        ]
+        for param_name, weight_name, shard_id, num_shard in stacked_params_mapping:
+            if weight_name in name:
+                return (
+                    name.replace(weight_name, param_name)[: -len(".weight")],
+                    num_shard,
+                )
+        return name[: -len(".weight")], 1
+
+    def get_num_params(self):
+        params_dict = dict(self.named_parameters())
+        return len(params_dict)
+
+    def load_weights(
+        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
+    ):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "c_fc_0", 0),
+            ("gate_up_proj", "c_fc_1", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+
+        def load_weights_per_param(name, loaded_weight):
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                return
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                return
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                return
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    return
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+        if name is None or loaded_weight is None:
+            for name, loaded_weight in weights:
+                name = name.replace("attn.attention", "self_attn")
+                load_weights_per_param(name, loaded_weight)
+        else:
+            name = name.replace("attn.attention", "self_attn")
+            load_weights_per_param(name, loaded_weight)
+
+
+EntryClass = ExaoneForCausalLM

sglang/srt/models/gemma.py

@@ -23,7 +23,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.activation import GeluAndMul
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
     QKVParallelLinear,
@@ -34,9 +33,11 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
 
@@ -60,7 +61,7 @@ class GemmaMLP(nn.Module):
             bias=False,
             quant_config=quant_config,
         )
-        self.act_fn = GeluAndMul()
+        self.act_fn = GeluAndMul("none")
 
     def forward(self, x):
         gate_up, _ = self.gate_up_proj(x)
@@ -287,6 +288,7 @@ class GemmaForCausalLM(nn.Module):
         self.quant_config = quant_config
         self.model = GemmaModel(config, quant_config=quant_config)
         self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
 
     @torch.no_grad()
     def forward(
@@ -297,9 +299,11 @@ class GemmaForCausalLM(nn.Module):
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        return self.logits_processor(
+        logits_output = self.logits_processor(
             input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
         )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return (sample_output, logits_output)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [