Add Muse Glimmer model support (#51655)

Signed-off-by: Tiezhen Wang <tiezhen@inferact.ai>
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Signed-off-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: Beto de Paola <betodepaola@meta.com>
Co-authored-by: Giancarlo Delfin <gdelfin@inferact.ai>
Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: zixi-qi <zixi@inferact.ai>
This commit is contained in:
Tiezhen WANG
2026-08-14 12:28:56 +08:00
committed by GitHub
co-authored by Beto de Paola Giancarlo Delfin Jee Jee Li Claude Opus 5 Jee Jee Li zixi-qi
parent 59d1af5698
commit 6adad08767
21 changed files with 4333 additions and 16 deletions
+18
View File
@@ -1192,6 +1192,12 @@ _MULTIMODAL_EXAMPLE_MODELS = {
# required by current PrefixLM implementation
max_num_batched_tokens=31872,
),
"MuseGlimmerForConditionalGeneration": _HfExamplesInfo(
"meta-models/Muse-Glimmer-30B",
),
"MuseGlimmerForCausalLM": _HfExamplesInfo(
"meta-models/Muse-Glimmer-30B",
),
"NVLM_D": _HfExamplesInfo("nvidia/NVLM-D-72B", trust_remote_code=True),
"Llama_Nemotron_Nano_VL": _HfExamplesInfo(
"nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
@@ -1489,6 +1495,18 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
max_num_seqs=32,
min_transformers_version="4.56.3", # Required for Qwen3Next
),
"MuseGlimmerAssistantModel": _HfExamplesInfo(
"meta-models/Muse-Glimmer-30B",
speculative_model="meta-models/Muse-Glimmer-30B-assistant",
max_model_len=8192, # Reduce max len to ensure test runs in low-VRAM CI env
max_num_seqs=32,
),
"DFlashMuseGlimmerAssistantModel": _HfExamplesInfo(
"meta-models/Muse-Glimmer-30B",
speculative_model="meta-models/Muse-Glimmer-30B-assistant",
max_model_len=8192, # Reduce max len to ensure test runs in low-VRAM CI env
max_num_seqs=32,
),
# [DSpark]
"DSparkDraftModel": _HfExamplesInfo(
"deepseek-ai/DeepSeek-V4-Pro-DSpark",
+7 -2
View File
@@ -564,15 +564,20 @@ def dummy_hf_overrides(
)
if hasattr(hf_config, "vision_config"):
hf_config.vision_config.update(
vision_config = hf_config.vision_config
vision_config.update(
{
"num_layers": 1,
"num_hidden_layers": 1,
}
)
# Keep per-layer metadata consistent with the reduced layer count.
if layer_types := getattr(vision_config, "layer_types", None):
vision_config.update({"layer_types": layer_types[:1]})
if model_arch in ("Moondream3ForCausalLM", "HfMoondream"):
hf_config.vision_config.update({"enc_n_layers": 1})
vision_config.update({"enc_n_layers": 1})
# e.g.: ibm-granite/granite-speech-3.3-2b
if hasattr(hf_config, "encoder_config"):