Export DSpark block drafters for Qwen 3.5 and 3.8

src/python/py/models/README.md

@@ -25,6 +25,7 @@ This folder contains the model builder for quickly creating optimized and quanti
     - [Include Auxiliary Hidden States Output](#include-auxiliary-hidden-states-output)
     - [Build with Paged Attention](#build-with-paged-attention)
     - [Build a DFlash 2 Block Drafter](#build-a-dflash-2-block-drafter)
+    - [Build a DSpark Block Drafter](#build-a-dspark-block-drafter)
     - [Disable Windowed KV Cache](#disable-windowed-kv-cache)
     - [Enable Shared Embeddings](#enable-shared-embeddings)
     - [Enable CUDA Graph Capture](#enable-cuda-graph-capture)
@@ -324,6 +325,20 @@ python -m onnxruntime_genai.models.builder -i path_to_target_model -o path_to_ou
 python builder.py -i path_to_target_model -o path_to_output_folder -p fp16 -e cuda -c cache_dir_for_hf_files --extra_options use_paged_attention=true aux_hidden_state_layers=1,11,21 dflash2_path=path_to_dflash2_checkpoint dflash2_num_draft_tokens=4
 ```
 
+#### Build a DSpark Block Drafter
+
+Set `dspark_path` to a DSpark checkpoint to export an auxiliary `dspark.onnx` block drafter beside a Qwen3.5 or Qwen3.8 target model. The target must use paged attention. SpecForge identifies the target layers whose outputs are tapped, while `aux_hidden_state_layers` identifies residual streams entering layers, so each configured auxiliary layer must be one greater than the corresponding `target_layer_ids` entry in the DSpark checkpoint. The drafter reuses the target's embedding and LM-head initializers. `dspark_path` and `dflash2_path` are mutually exclusive.
+
+`dspark_num_draft_tokens` optionally overrides how many tokens the drafter proposes per step and must be a positive integer. `dspark_top_k` controls how many candidates the lattice keeps per block slot; it defaults to `16` and must be a positive integer no greater than the drafter vocabulary size.
+
+```bash
+# From wheel:
+python -m onnxruntime_genai.models.builder -i path_to_target_model -o path_to_output_folder -p fp16 -e cuda -c cache_dir_for_hf_files --extra_options use_paged_attention=true aux_hidden_state_layers=1,11,21 dspark_path=path_to_dspark_checkpoint dspark_num_draft_tokens=4 dspark_top_k=16
+
+# From source:
+python builder.py -i path_to_target_model -o path_to_output_folder -p fp16 -e cuda -c cache_dir_for_hf_files --extra_options use_paged_attention=true aux_hidden_state_layers=1,11,21 dspark_path=path_to_dspark_checkpoint dspark_num_draft_tokens=4 dspark_top_k=16
+```
+
 #### Disable Windowed KV Cache
 
 By default, sliding-window layers use a reduced KV cache on supported execution providers. With paged attention, eligible local layers use a ring of blocks when the exported model also contains at least one full-context layer. Set `windowed_kv_cache=false` to give every layer a full-length KV cache, which is useful for performance comparisons or compatibility testing. The option defaults to `true` and applies to both paged and non-paged models.

src/python/py/models/builder.py

@@ -757,6 +757,15 @@ def get_args():
                     aux_hidden_state_layers to match the drafter's `target_layer_ids`. Default is unset (disabled).
                 dflash2_num_draft_tokens = Override the number of draft tokens the DFlash 2 block
                     drafter proposes per step. Must be a positive integer. Default is taken from the draft checkpoint.
+                dspark_path = Path to a DSpark draft checkpoint. Exports an auxiliary `dspark.onnx`
+                    block drafter beside the target model and adds a `dspark` section to
+                    genai_config.json. Mutually exclusive with dflash2_path. Requires
+                    use_paged_attention=true. SpecForge taps each target layer's output, so
+                    aux_hidden_state_layers must be the drafter's `target_layer_ids` each plus one.
+                dspark_num_draft_tokens = Override the number of draft tokens the DSpark block
+                    drafter proposes per step. Must be a positive integer. Default is taken from the draft checkpoint.
+                dspark_top_k = Candidates the DSpark lattice keeps per block slot. Must be a positive integer no
+                    greater than the drafter vocabulary size. Default is 16.
                 mtp_quant_config = JSON object/file: Configure MTP I/O, dense weights, MoE, and runtime using the
                     structured QuantConfig schema independently from the main model.
                 linear_attn_op = linear_attention/gated_delta_net: Select the recurrent operator for non-paged