[Spec Decode] DSpark (#46995)

Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Signed-off-by: Giancarlo Delfin <gdelfin@inferact.ai>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Giancarlo Delfin <gdelfin@inferact.ai>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Benjamin Chislett
2026-07-01 08:30:24 -07:00
committed by GitHub
co-authored by Giancarlo Delfin mgoin
parent 63fcce4de1
commit f5a8d73377
24 changed files with 1821 additions and 94 deletions
+13
View File
@@ -1407,6 +1407,19 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
max_num_seqs=32,
min_transformers_version="4.56.3", # Required for Qwen3Next
),
# [DSpark]
"DSparkDraftModel": _HfExamplesInfo(
"deepseek-ai/DeepSeek-V4-Pro-DSpark",
speculative_model="deepseek-ai/DeepSeek-V4-Pro-DSpark", # draft in mtp.*
is_available_online=False,
use_original_num_layers=True, # DSpark has >1 draft block
),
"Qwen3DSparkModel": _HfExamplesInfo(
"Qwen/Qwen3-8B",
speculative_model="deepseek-ai/dspark_qwen3_8b_block7",
is_available_online=False,
use_original_num_layers=True, # DSpark backbone requires all layers
),
# [Eagle]
"EagleCohereForCausalLM": _HfExamplesInfo(
"/host/engines/cohere-moe",
+4
View File
@@ -48,6 +48,10 @@ def test_registry_imports(model_arch):
"(see #41376)"
)
# DSpark draft model is NVIDIA-only; class is stubbed to None on ROCm/XPU.
if model_arch == "DSparkDraftModel" and not current_platform.is_cuda():
pytest.skip("DSparkDraftModel is only supported on CUDA")
# Ensure all model classes can be imported successfully
model_cls = ModelRegistry._try_load_model_cls(model_arch)
assert model_cls is not None