A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 15:15 UTC
[Model] Support InternS2 Preview (#42705)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: zxy <46674730+CUHKSZzxy@users.noreply.github.com>
Isotr0py committed
May 15, 2026 at 12:30 UTC
faa4b76afae3660efc8a5d48a8c7c2d37e0ffb05
6 files changed
+60
docs/models/supported_models.md
+1
@@ -575,6 +575,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
575
| `IsaacForConditionalGeneration` | Isaac | T + I<sup>+</sup> | `PerceptronAI/Isaac-0.1` | ✅︎ | ✅︎ |
576
| `InternS1ForConditionalGeneration` | Intern-S1 | T + I<sup>E+</sup> + V<sup>E+</sup> | `internlm/Intern-S1`, `internlm/Intern-S1-mini`, etc. | ✅︎ | ✅︎ |
577
| `InternS1ProForConditionalGeneration` | Intern-S1-Pro | T + I<sup>E+</sup> + V<sup>E+</sup> | `internlm/Intern-S1-Pro`, etc. | ✅︎ | ✅︎ |
578
+| `InternS2PreviewForConditionalGeneration` | Intern-S2-Preview | T + I<sup>E+</sup> + V<sup>E+</sup> | `internlm/Intern-S2-Preview`, etc. | ✅︎ | ✅︎ |
579
| `InternVLChatModel` | InternVL 3.5, InternVL 3.0, InternVideo 2.5, InternVL 2.5, Mono-InternVL, InternVL 2.0 | T + I<sup>E+</sup> + (V<sup>E+</sup>) | `OpenGVLab/InternVL3_5-14B`, `OpenGVLab/InternVL3-9B`, `OpenGVLab/InternVideo2_5_Chat_8B`, `OpenGVLab/InternVL2_5-4B`, `OpenGVLab/Mono-InternVL-2B`, `OpenGVLab/InternVL2-4B`, etc. | ✅︎ | ✅︎ |
580
| `InternVLForConditionalGeneration` | InternVL 3.0 (HF format) | T + I<sup>E+</sup> + V<sup>E+</sup> | `OpenGVLab/InternVL3-1B-hf`, etc. | ✅︎ | ✅︎ |
581
| `KananaVForConditionalGeneration` | Kanana-V | T + I<sup>+</sup> | `kakaocorp/kanana-1.5-v-3b-instruct`, etc. | | ✅︎ |
tests/models/registry.py
+5
@@ -988,6 +988,11 @@ _MULTIMODAL_EXAMPLE_MODELS = {
988
"internlm/Intern-S1-Pro",
989
trust_remote_code=True,
990
),
991
+ "InternS2PreviewForConditionalGeneration": _HfExamplesInfo(
992
+ "internlm/Intern-S2-Preview",
993
+ trust_remote_code=True,
994
+ is_available_online=False,
995
+ ),
996
"InternVLChatModel": _HfExamplesInfo(
997
"OpenGVLab/InternVL2-1B",
998
extras={
vllm/config/speculative.py
+11
@@ -467,6 +467,17 @@ class SpeculativeConfig:
467
"architectures": ["Qwen3_5MoeMTP" if is_moe else "Qwen3_5MTP"],
468
}
469
)
470
+ if hf_config.model_type == "intern_s2_preview":
471
+ text_config = getattr(hf_config, "text_config", None)
472
+ is_moe = getattr(text_config, "model_type", None) == "qwen3_5_moe_text"
473
+ hf_config.model_type = "qwen3_5_mtp"
474
+ n_predict = getattr(text_config, "mtp_num_hidden_layers", None)
475
+ hf_config.update(
476
+ {
477
+ "n_predict": n_predict,
478
+ "architectures": ["Qwen3_5MoeMTP" if is_moe else "Qwen3_5MTP"],
479
+ }
480
+ )
481
if hf_config.model_type == "longcat_flash":
482
hf_config.model_type = "longcat_flash_mtp"
483
n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
vllm/model_executor/models/interns2_preview.py
new
+38
@@ -0,0 +1,38 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+from collections.abc import Iterable
4
+
5
+import torch
6
+from transformers import AutoProcessor
7
+
8
+from vllm.multimodal import MULTIMODAL_REGISTRY
9
+
10
+from .qwen3_5 import Qwen3_5MoeForConditionalGeneration
11
+from .qwen3_vl import (
12
+ Qwen3VLDummyInputsBuilder,
13
+ Qwen3VLMultiModalProcessor,
14
+ Qwen3VLProcessingInfo,
15
+)
16
+from .utils import AutoWeightsLoader
17
+
18
+
19
+class InternS2PreviewProcessingInfo(Qwen3VLProcessingInfo):
20
+ def get_hf_config(self):
21
+ return self.ctx.get_hf_config()
22
+
23
+ def get_hf_processor(self, **kwargs: object) -> AutoProcessor:
24
+ return self.ctx.get_hf_processor(**kwargs)
25
+
26
+
27
+@MULTIMODAL_REGISTRY.register_processor(
28
+ Qwen3VLMultiModalProcessor,
29
+ info=InternS2PreviewProcessingInfo,
30
+ dummy_inputs=Qwen3VLDummyInputsBuilder,
31
+)
32
+class InternS2PreviewForConditionalGeneration(Qwen3_5MoeForConditionalGeneration):
33
+ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
34
+ loader = AutoWeightsLoader(
35
+ self,
36
+ skip_prefixes=["mtp.", "model.time_series.", "time_series."],
37
+ )
38
+ return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
vllm/model_executor/models/registry.py
+4
@@ -435,6 +435,10 @@ _MULTIMODAL_MODELS = {
435
"interns1_pro",
436
"InternS1ProForConditionalGeneration",
437
),
438
+ "InternS2PreviewForConditionalGeneration": (
439
+ "interns2_preview",
440
+ "InternS2PreviewForConditionalGeneration",
441
+ ),
442
"Idefics3ForConditionalGeneration": (
443
"idefics3",
444
"Idefics3ForConditionalGeneration",
vllm/v1/spec_decode/llm_base_proposer.py
+1
@@ -1218,6 +1218,7 @@ class SpecDecodeBaseProposer:
1218
"Exaone4_5_ForConditionalGeneration",
1219
"GlmOcrForConditionalGeneration",
1220
"HunYuanVLForConditionalGeneration",
1221
+ "InternS2PreviewForConditionalGeneration",
1222
"MiMoV2OmniForCausalLM",
1223
"Qwen2_5_VLForConditionalGeneration",
1224
"Qwen3_5ForConditionalGeneration",