[None][feat] Add DeepSeek DSpark speculative decoding

tensorrt_llm/_torch/attention_backend/sparse/deepseek_v4/deepseek_v4.py

@@ -529,6 +529,10 @@ def prepare_for_indexer_k_cache(self):
             self.host_indexer_k_cache_block_offsets[: self.num_seqs],
             non_blocking=True,
         )
+        # Columns beyond each sequence's allocated indexer blocks contain BAD_PAGE_INDEX (-1).
+        # CUDA-graph padded token slots may still compute scatter addresses from those columns
+        # before being ignored, so map them to block 0, matching the base DSA metadata path.
+        self.indexer_k_cache_block_offsets.clamp_(min=0)
 
     def prepare_for_block_tables(self):
         """Prepare block tables for sliding-window and compressed attention."""

tensorrt_llm/_torch/model_config.py

@@ -956,9 +956,14 @@ def update_sparse_attention_indexer_config(pretrained_config, kwargs):
                         pretrained_config, 'compress_ratios', None)
                     num_base_layers = pretrained_config.num_hidden_layers
                     spec_config = kwargs.get('spec_config', None)
-                    if (spec_config is not None
-                            and getattr(spec_config, 'num_nextn_predict_layers',
-                                        None) is None):
+                    # ``num_nextn_predict_layers`` is MTP-specific (only read on
+                    # the is_mtp_one_model path). Only set it on configs that
+                    # actually declare the field; other DeepSeek-V4 spec modes
+                    # (e.g. DSpark, which carries its own draft stage count) do
+                    # not, and a blind setattr would fail pydantic validation.
+                    if (spec_config is not None and 'num_nextn_predict_layers'
+                            in type(spec_config).model_fields
+                            and spec_config.num_nextn_predict_layers is None):
                         spec_config.num_nextn_predict_layers = getattr(
                             pretrained_config, 'num_nextn_predict_layers', 1)
                     mtp_enabled = (spec_config is not None and
@@ -990,6 +995,22 @@ def update_sparse_attention_indexer_config(pretrained_config, kwargs):
                     if window_size is None:
                         window_size = pretrained_config.sliding_window
 
+                    # DeepSeek-V4 needs explicit per-layer compress ratios. They
+                    # must come from the checkpoint config or a user override; we
+                    # intentionally do not synthesize a default list (it would
+                    # silently change sparse-attention semantics). Fail fast with
+                    # an actionable message instead of letting the normalization
+                    # below raise an opaque TypeError on None.
+                    if compress_ratios is None:
+                        raise ValueError(
+                            "DeepSeek-V4 requires per-layer `compress_ratios`, "
+                            "but none were found in the checkpoint config and "
+                            "none were provided via `sparse_attention_config`. "
+                            "Set `compress_ratios` in the model's config.json, or "
+                            "pass `sparse_attention_config="
+                            "DeepSeekV4SparseAttentionConfig(compress_ratios=[...])`"
+                            " in --extra_llm_api_options.")
+
                     # Normalize checkpoint-facing ratio 0 (SWA-only/uncompressed)
                     # to 1 internally so cache allocation math works. The
                     # external config keeps the original semantics.

tensorrt_llm/_torch/models/dspark/__init__.py

@@ -0,0 +1,16 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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.
+
+"""DSpark draft-model components."""