[None][feat] Support Step-3.7-Flash model

docs/source/features/speculative-decoding.md

@@ -103,7 +103,7 @@ llm = LLM("/path/to/target_model", speculative_config=speculative_config, disabl
 
 ### MTP
 
-MTP is currently only supported by Deepseek. MTP can be tuned with the following configuration options:
+MTP is supported by DeepSeek models and other architectures that ship native MTP modules (including Step-3.x). MTP can be tuned with the following configuration options:
 
 * `max_draft_len`: Maximum draft candidate length.
 * `num_nextn_predict_layers`: Number of MTP modules to use. Currently must match `max_draft_len`.

docs/source/models/supported-models.md

@@ -50,6 +50,7 @@ The following is a table of supported models for the PyTorch backend:
 | `SeedOssForCausalLM` [^5]            | Seed OSS, Seed-Coder               | `ByteDance-Seed/Seed-OSS-36B-Instruct`       |
 | `SkyworkR1V2ForConditionalGeneration` [^5] | Skywork R1V2, Skywork SWE    | `Skywork/Skywork-R1V2-38B`                   |
 | `SmolLM3ForCausalLM` [^5]            | SmolLM3                            | `HuggingFaceTB/SmolLM3-3B`                   |
+| `Step3p7ForConditionalGeneration` [^8]| Step-3.7-Flash                    | `stepfun-ai/Step-3.7-Flash`                  |
 
 
 ## Model-Feature Support Matrix (Key Models)
@@ -69,13 +70,15 @@ Note: Support for other models may vary. Features marked "N/A" are not applicabl
 | `Glm4MoeLiteForCausalLM` [^5]    | Yes               | Yes        | Untested                   | Untested              | Yes             | No  | No               | No                | No     | Yes           | Untested         | Untested       | N/A                      | Untested              | Untested        |
 | `NemotronHForCausalLM` (Super)   | Yes               | Yes        | Untested                   | Untested              | Yes             | Yes | No               | No                | No     | Yes           | Yes              | Untested       | N/A                      | Untested              | Untested        |
 | `Gemma4ForConditionalGeneration` | Untested          | Yes        | Untested                   | No                    | Yes             | No  | No               | No                | No     | Yes           | Untested         | No             | Yes                      | Untested              | Untested        |
+| `Step3p7ForConditionalGeneration`| Yes               | Yes        | Yes                        | Untested              | Untested        | Yes | No               | No                | No     | Yes           | Untested         | Untested       | Yes                      | Untested              | Untested        |
 
 [^1]: Chunked Prefill for MLA can only be enabled on SM100/SM103.
 [^2]: KV cache reuse for MLA can only be enabled on SM90/SM100/SM103 and in BF16/FP8 KV cache dtype.
 [^3]: Qwen3-Next-80B-A3B exhibits relatively low accuracy on the SciCode-AA-v2 benchmark.
 [^5]: Supported via the [AutoDeploy](../features/auto_deploy/auto-deploy.md) backend. See [AD Configs](../../../examples/auto_deploy/model_registry/configs).
 [^6]: Also supports text-only inference via the [AutoDeploy](../features/auto_deploy/auto-deploy.md) backend.
 [^7]: Text-only support via the [AutoDeploy](../features/auto_deploy/auto-deploy.md) backend.
+[^8]: Supports text and image inputs. The vision tower runs in BF16 even when the text decoder is quantized (FP8 block-scale or NVFP4). The text decoder is also usable standalone (text-only) via the `Step3p5ForCausalLM` architecture.
 [^9]: Audio modality only supported on E2B/E4B variants.
 
 # Multimodal Feature Support Matrix (PyTorch Backend)
@@ -96,6 +99,7 @@ Note: Support for other models may vary. Features marked "N/A" are not applicabl
 | `Qwen2_5_VLForConditionalGeneration` | Yes               | Yes        | Yes             | Yes           | Yes              | Yes            | Yes                   | Yes                       | L + I + V |
 | `Qwen3VLForConditionalGeneration`    | Yes               | Yes        | Yes             | Yes           | Yes              | Yes            | Yes                   | Yes                       | L + I + V |
 | `Qwen3VLMoeForConditionalGeneration` | Yes               | Yes        | Yes             | Yes           | Yes              | Yes            | Yes                   | Yes                       | L + I + V |
+| `Step3p7ForConditionalGeneration`    | Yes               | Yes        | Untested        | Yes           | Untested         | Untested       | Untested              | Untested                  | L + I     |
 
 Note:
 - L: Language

tensorrt_llm/_torch/models/__init__.py

@@ -46,6 +46,8 @@
 from .modeling_seedoss import SeedOssForCausalLM
 from .modeling_siglip import SiglipVisionModel
 from .modeling_starcoder2 import Starcoder2ForCausalLM
+from .modeling_step3p7 import Step3p7ForCausalLM
+from .modeling_step3p7vl import Step3p7VLForConditionalGeneration
 from .modeling_utils import get_model_architecture
 from .modeling_vila import VilaModel
 
@@ -98,6 +100,8 @@
     "Qwen3VLModel",
     "MiniMaxM2ForCausalLM",
     "Cohere2ForCausalLM",
+    "Step3p7ForCausalLM",
+    "Step3p7VLForConditionalGeneration",
 ]
 
 if transformers.__version__ >= "4.45.1":

tensorrt_llm/_torch/models/modeling_speculative.py

@@ -1423,6 +1423,9 @@ def __init__(
             case "qwen3_next" | "qwen3_5_text" | "qwen3_5_moe_text":
                 from .modeling_qwen3_next import Qwen3NextMTP
                 mtp_layer = Qwen3NextMTP
+            case "step3p7" | "step3p5":
+                from .modeling_step3p7 import Step3p7MTP
+                mtp_layer = Step3p7MTP
             case _:
                 raise ValueError(
                     f"Model type {model_type} not supported for MTP")