sglang 0.3.5.post2__py3-none-any.whl → 0.3.6.post1__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- sglang/__init__.py +2 -2
- sglang/api.py +2 -2
- sglang/bench_latency.py +1 -553
- sglang/bench_offline_throughput.py +48 -20
- sglang/bench_one_batch.py +472 -0
- sglang/{bench_server_latency.py → bench_one_batch_server.py} +3 -3
- sglang/bench_serving.py +125 -6
- sglang/check_env.py +3 -6
- sglang/lang/backend/base_backend.py +1 -1
- sglang/lang/backend/runtime_endpoint.py +2 -2
- sglang/srt/configs/model_config.py +13 -14
- sglang/srt/constrained/__init__.py +13 -14
- sglang/srt/constrained/base_grammar_backend.py +13 -15
- sglang/srt/constrained/outlines_backend.py +28 -17
- sglang/srt/constrained/outlines_jump_forward.py +13 -15
- sglang/srt/constrained/xgrammar_backend.py +47 -58
- sglang/srt/conversation.py +13 -15
- sglang/srt/hf_transformers_utils.py +13 -15
- sglang/srt/layers/activation.py +16 -13
- sglang/srt/layers/attention/flashinfer_backend.py +106 -54
- sglang/srt/layers/attention/triton_backend.py +9 -7
- sglang/srt/layers/attention/triton_ops/decode_attention.py +51 -55
- sglang/srt/layers/attention/triton_ops/extend_attention.py +16 -16
- sglang/srt/layers/attention/triton_ops/prefill_attention.py +13 -15
- sglang/srt/layers/custom_op_util.py +25 -0
- sglang/srt/layers/fused_moe_grok/__init__.py +1 -0
- sglang/srt/layers/{fused_moe → fused_moe_grok}/fused_moe.py +11 -4
- sglang/srt/layers/{fused_moe → fused_moe_grok}/layer.py +4 -9
- sglang/srt/layers/{fused_moe/patch.py → fused_moe_patch.py} +5 -0
- sglang/srt/layers/fused_moe_triton/__init__.py +44 -0
- sglang/srt/layers/fused_moe_triton/fused_moe.py +861 -0
- sglang/srt/layers/fused_moe_triton/layer.py +633 -0
- sglang/srt/layers/layernorm.py +17 -15
- sglang/srt/layers/logits_processor.py +23 -25
- sglang/srt/layers/quantization/__init__.py +77 -17
- sglang/srt/layers/radix_attention.py +13 -15
- sglang/srt/layers/rotary_embedding.py +13 -13
- sglang/srt/layers/sampler.py +4 -8
- sglang/srt/layers/torchao_utils.py +2 -0
- sglang/srt/lora/lora.py +13 -14
- sglang/srt/lora/lora_config.py +13 -14
- sglang/srt/lora/lora_manager.py +22 -24
- sglang/srt/managers/data_parallel_controller.py +98 -27
- sglang/srt/managers/detokenizer_manager.py +13 -15
- sglang/srt/managers/io_struct.py +63 -21
- sglang/srt/managers/schedule_batch.py +154 -59
- sglang/srt/managers/schedule_policy.py +18 -16
- sglang/srt/managers/scheduler.py +278 -109
- sglang/srt/managers/session_controller.py +61 -0
- sglang/srt/managers/tokenizer_manager.py +63 -18
- sglang/srt/managers/tp_worker.py +25 -16
- sglang/srt/managers/tp_worker_overlap_thread.py +62 -67
- sglang/srt/metrics/collector.py +13 -15
- sglang/srt/metrics/func_timer.py +13 -15
- sglang/srt/mm_utils.py +13 -14
- sglang/srt/model_executor/cuda_graph_runner.py +63 -25
- sglang/srt/model_executor/forward_batch_info.py +128 -32
- sglang/srt/model_executor/model_runner.py +132 -64
- sglang/srt/model_parallel.py +98 -0
- sglang/srt/models/chatglm.py +15 -16
- sglang/srt/models/commandr.py +15 -16
- sglang/srt/models/dbrx.py +15 -16
- sglang/srt/models/deepseek.py +15 -15
- sglang/srt/models/deepseek_v2.py +162 -59
- sglang/srt/models/exaone.py +14 -15
- sglang/srt/models/gemma.py +14 -14
- sglang/srt/models/gemma2.py +31 -25
- sglang/srt/models/gemma2_reward.py +13 -14
- sglang/srt/models/gpt_bigcode.py +14 -14
- sglang/srt/models/grok.py +15 -15
- sglang/srt/models/internlm2.py +13 -15
- sglang/srt/models/internlm2_reward.py +13 -14
- sglang/srt/models/llama.py +21 -21
- sglang/srt/models/llama_classification.py +13 -14
- sglang/srt/models/llama_reward.py +13 -14
- sglang/srt/models/llava.py +14 -16
- sglang/srt/models/llavavid.py +14 -16
- sglang/srt/models/minicpm.py +13 -15
- sglang/srt/models/minicpm3.py +13 -15
- sglang/srt/models/mistral.py +13 -15
- sglang/srt/models/mixtral.py +15 -15
- sglang/srt/models/mixtral_quant.py +14 -14
- sglang/srt/models/olmo.py +22 -20
- sglang/srt/models/olmoe.py +23 -20
- sglang/srt/models/phi3_small.py +447 -0
- sglang/srt/models/qwen.py +14 -14
- sglang/srt/models/qwen2.py +22 -19
- sglang/srt/models/qwen2_moe.py +17 -18
- sglang/srt/models/qwen2_vl.py +13 -6
- sglang/srt/models/stablelm.py +18 -16
- sglang/srt/models/torch_native_llama.py +107 -93
- sglang/srt/models/xverse.py +13 -14
- sglang/srt/models/xverse_moe.py +15 -16
- sglang/srt/models/yivl.py +13 -15
- sglang/srt/openai_api/adapter.py +19 -17
- sglang/srt/openai_api/protocol.py +14 -16
- sglang/srt/sampling/penaltylib/orchestrator.py +49 -79
- sglang/srt/sampling/penaltylib/penalizers/frequency_penalty.py +3 -8
- sglang/srt/sampling/penaltylib/penalizers/min_new_tokens.py +3 -9
- sglang/srt/sampling/penaltylib/penalizers/presence_penalty.py +3 -8
- sglang/srt/sampling/penaltylib/penalizers/repetition_penalty.py +3 -8
- sglang/srt/sampling/sampling_batch_info.py +61 -57
- sglang/srt/sampling/sampling_params.py +14 -16
- sglang/srt/server.py +86 -35
- sglang/srt/server_args.py +96 -80
- sglang/srt/utils.py +266 -68
- sglang/test/few_shot_gsm8k.py +8 -4
- sglang/test/runners.py +38 -20
- sglang/test/srt/sampling/penaltylib/utils.py +23 -21
- sglang/test/test_utils.py +31 -20
- sglang/version.py +1 -1
- {sglang-0.3.5.post2.dist-info → sglang-0.3.6.post1.dist-info}/LICENSE +1 -1
- {sglang-0.3.5.post2.dist-info → sglang-0.3.6.post1.dist-info}/METADATA +66 -57
- sglang-0.3.6.post1.dist-info/RECORD +164 -0
- {sglang-0.3.5.post2.dist-info → sglang-0.3.6.post1.dist-info}/WHEEL +1 -1
- sglang/srt/layers/fused_moe/__init__.py +0 -1
- sglang-0.3.5.post2.dist-info/RECORD +0 -156
- {sglang-0.3.5.post2.dist-info → sglang-0.3.6.post1.dist-info}/top_level.txt +0 -0
sglang/srt/models/internlm2.py
CHANGED
@@ -1,19 +1,17 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
6
14
|
|
7
|
-
http://www.apache.org/licenses/LICENSE-2.0
|
8
|
-
|
9
|
-
Unless required by applicable law or agreed to in writing, software
|
10
|
-
distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
-
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
-
See the License for the specific language governing permissions and
|
13
|
-
limitations under the License.
|
14
|
-
"""
|
15
|
-
|
16
|
-
# -*- coding: utf-8 -*-
|
17
15
|
# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/7f62077af5159c625fe3ad1c812e6c1a2b93ba3b/vllm/model_executor/models/internlm2.py
|
18
16
|
|
19
17
|
from typing import Any, Dict, Iterable, Optional, Tuple
|
@@ -1,17 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
from typing import Iterable, Optional, Tuple
|
17
16
|
|
sglang/srt/models/llama.py
CHANGED
@@ -1,17 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
# Adapted from
|
17
16
|
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/llama.py#L1
|
@@ -44,6 +43,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
|
|
44
43
|
)
|
45
44
|
from sglang.srt.managers.schedule_batch import global_server_args_dict
|
46
45
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
46
|
+
from sglang.srt.utils import make_layers
|
47
47
|
|
48
48
|
|
49
49
|
class LlamaMLP(nn.Module):
|
@@ -256,14 +256,14 @@ class LlamaModel(nn.Module):
|
|
256
256
|
config.vocab_size,
|
257
257
|
config.hidden_size,
|
258
258
|
)
|
259
|
-
self.layers =
|
260
|
-
|
261
|
-
|
262
|
-
|
263
|
-
|
264
|
-
|
265
|
-
]
|
259
|
+
self.layers = make_layers(
|
260
|
+
config.num_hidden_layers,
|
261
|
+
lambda idx, prefix: LlamaDecoderLayer(
|
262
|
+
config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
|
263
|
+
),
|
264
|
+
prefix="model.layers",
|
266
265
|
)
|
266
|
+
|
267
267
|
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
268
268
|
|
269
269
|
def forward(
|
@@ -1,17 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
from typing import Iterable, Optional, Tuple
|
17
16
|
|
@@ -1,17 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
from typing import Iterable, Optional, Tuple
|
17
16
|
|
sglang/srt/models/llava.py
CHANGED
@@ -1,18 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
15
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
16
14
|
"""Inference-only LLaVa model compatible with HuggingFace weights."""
|
17
15
|
|
18
16
|
import math
|
@@ -345,7 +343,7 @@ class LlavaBaseForCausalLM(nn.Module):
|
|
345
343
|
|
346
344
|
# Fill in the placeholder for the image
|
347
345
|
extend_start_loc_cpu = forward_batch.extend_start_loc.cpu().numpy()
|
348
|
-
prefix_lens_cpu = forward_batch.
|
346
|
+
prefix_lens_cpu = forward_batch.extend_prefix_lens_cpu
|
349
347
|
pt = 0
|
350
348
|
for i in range(bs):
|
351
349
|
if not need_vision[i]:
|
sglang/srt/models/llavavid.py
CHANGED
@@ -1,18 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
15
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
16
14
|
"""Inference-only LLaVa video model compatible with HuggingFace weights."""
|
17
15
|
|
18
16
|
from typing import Iterable, List, Optional, Tuple
|
@@ -169,7 +167,7 @@ class LlavaVidForCausalLM(nn.Module):
|
|
169
167
|
|
170
168
|
# Fill in the placeholder for the image
|
171
169
|
extend_start_loc_cpu = forward_batch.extend_start_loc.cpu().numpy()
|
172
|
-
prefix_lens_cpu = forward_batch.
|
170
|
+
prefix_lens_cpu = forward_batch.extend_prefix_lens_cpu
|
173
171
|
pt = 0
|
174
172
|
for i in range(bs):
|
175
173
|
if not need_vision[i]:
|
sglang/srt/models/minicpm.py
CHANGED
@@ -1,18 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
15
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
16
14
|
"""Inference-only MiniCPM model compatible with HuggingFace weights."""
|
17
15
|
|
18
16
|
import math
|
sglang/srt/models/minicpm3.py
CHANGED
@@ -1,18 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
15
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
16
14
|
"""Inference-only MiniCPM3 model compatible with HuggingFace weights."""
|
17
15
|
|
18
16
|
import math
|
sglang/srt/models/mistral.py
CHANGED
@@ -1,18 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
15
|
-
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
16
14
|
"""Inference-only Mistral model."""
|
17
15
|
|
18
16
|
from sglang.srt.models.llama import LlamaForCausalLM
|
sglang/srt/models/mixtral.py
CHANGED
@@ -1,31 +1,31 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
# Adapted from
|
17
16
|
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/mixtral.py#L1
|
18
17
|
"""Inference-only Mixtral model."""
|
18
|
+
|
19
19
|
from typing import Iterable, Optional, Tuple
|
20
20
|
|
21
21
|
import torch
|
22
22
|
from torch import nn
|
23
23
|
from transformers import MixtralConfig
|
24
24
|
from vllm.distributed import get_tensor_model_parallel_world_size
|
25
|
-
from vllm.model_executor.layers.fused_moe import FusedMoE
|
26
25
|
from vllm.model_executor.layers.rotary_embedding import get_rope
|
27
26
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
28
27
|
|
28
|
+
from sglang.srt.layers.fused_moe_triton import FusedMoE
|
29
29
|
from sglang.srt.layers.layernorm import RMSNorm
|
30
30
|
from sglang.srt.layers.linear import (
|
31
31
|
QKVParallelLinear,
|
@@ -1,21 +1,21 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
# Adapted from
|
17
16
|
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/mixtral_quant.py#L1
|
18
17
|
"""Inference-only Mixtral model."""
|
18
|
+
|
19
19
|
from typing import Iterable, Optional, Tuple
|
20
20
|
|
21
21
|
import numpy as np
|
sglang/srt/models/olmo.py
CHANGED
@@ -1,17 +1,16 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
# Adapted from
|
17
16
|
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/olmo.py#L1
|
@@ -39,6 +38,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
|
|
39
38
|
VocabParallelEmbedding,
|
40
39
|
)
|
41
40
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
41
|
+
from sglang.srt.utils import make_layers
|
42
42
|
|
43
43
|
|
44
44
|
class OlmoAttention(nn.Module):
|
@@ -221,11 +221,13 @@ class OlmoModel(nn.Module):
|
|
221
221
|
self.embed_tokens = VocabParallelEmbedding(
|
222
222
|
config.vocab_size, config.hidden_size
|
223
223
|
)
|
224
|
-
self.layers =
|
225
|
-
|
226
|
-
|
227
|
-
|
228
|
-
|
224
|
+
self.layers = make_layers(
|
225
|
+
config.num_hidden_layers,
|
226
|
+
lambda idx, prefix: OlmoDecoderLayer(
|
227
|
+
layer_id=idx,
|
228
|
+
config=config,
|
229
|
+
quant_config=quant_config,
|
230
|
+
),
|
229
231
|
)
|
230
232
|
self.norm = nn.LayerNorm(
|
231
233
|
config.hidden_size, elementwise_affine=False, bias=False
|
@@ -250,7 +252,7 @@ class OlmoModel(nn.Module):
|
|
250
252
|
hidden_states = input_embeds
|
251
253
|
|
252
254
|
# Apply blocks one-by-one.
|
253
|
-
for
|
255
|
+
for layer_id, decoder_layer in enumerate(self.layers):
|
254
256
|
# shape: (batch_size, seq_len, d_model)
|
255
257
|
hidden_states = decoder_layer(
|
256
258
|
positions,
|
sglang/srt/models/olmoe.py
CHANGED
@@ -1,22 +1,22 @@
|
|
1
|
-
|
2
|
-
|
3
|
-
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
|
8
|
-
|
9
|
-
|
10
|
-
|
11
|
-
|
12
|
-
|
13
|
-
|
14
|
-
"""
|
1
|
+
# Copyright 2023-2024 SGLang Team
|
2
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3
|
+
# you may not use this file except in compliance with the License.
|
4
|
+
# You may obtain a copy of the License at
|
5
|
+
#
|
6
|
+
# http://www.apache.org/licenses/LICENSE-2.0
|
7
|
+
#
|
8
|
+
# Unless required by applicable law or agreed to in writing, software
|
9
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
10
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
11
|
+
# See the License for the specific language governing permissions and
|
12
|
+
# limitations under the License.
|
13
|
+
# ==============================================================================
|
15
14
|
|
16
15
|
# Adapted from:
|
17
16
|
# https://github.com/vllm-project/vllm/pull/7922
|
18
17
|
|
19
18
|
"""Inference-only OLMoE model compatible with HuggingFace weights."""
|
19
|
+
|
20
20
|
from typing import Any, Dict, Iterable, List, Optional, Tuple
|
21
21
|
|
22
22
|
import torch
|
@@ -27,7 +27,6 @@ from vllm.distributed import (
|
|
27
27
|
get_tensor_model_parallel_world_size,
|
28
28
|
tensor_model_parallel_all_reduce,
|
29
29
|
)
|
30
|
-
from vllm.model_executor.layers.fused_moe import FusedMoE
|
31
30
|
from vllm.model_executor.layers.linear import (
|
32
31
|
MergedColumnParallelLinear,
|
33
32
|
QKVParallelLinear,
|
@@ -39,6 +38,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
39
38
|
from vllm.utils import print_warning_once
|
40
39
|
|
41
40
|
from sglang.srt.layers.activation import SiluAndMul
|
41
|
+
from sglang.srt.layers.fused_moe_triton import FusedMoE
|
42
42
|
from sglang.srt.layers.layernorm import RMSNorm
|
43
43
|
from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
|
44
44
|
from sglang.srt.layers.quantization.base_config import QuantizationConfig
|
@@ -48,6 +48,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
|
|
48
48
|
VocabParallelEmbedding,
|
49
49
|
)
|
50
50
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
51
|
+
from sglang.srt.utils import make_layers
|
51
52
|
|
52
53
|
|
53
54
|
class OlmoeMoE(nn.Module):
|
@@ -261,11 +262,13 @@ class OlmoeModel(nn.Module):
|
|
261
262
|
config.vocab_size,
|
262
263
|
config.hidden_size,
|
263
264
|
)
|
264
|
-
self.layers =
|
265
|
-
|
266
|
-
|
267
|
-
|
268
|
-
|
265
|
+
self.layers = make_layers(
|
266
|
+
config.num_hidden_layers,
|
267
|
+
lambda idx, prefix: OlmoeDecoderLayer(
|
268
|
+
config=config,
|
269
|
+
quant_config=quant_config,
|
270
|
+
layer_id=idx,
|
271
|
+
),
|
269
272
|
)
|
270
273
|
self.norm = RMSNorm(config.hidden_size, eps=1e-5)
|
271
274
|
|