[Mamba] Add FlashInfer ReplaySSM support for MTP (#52928)

Signed-off-by: Andrii Skliar <askliar@nvidia.com>
Co-authored-by: Andrii Skliar <askliar@nvidia.com>
This commit is contained in:
Andrii Skliar
2026-09-30 21:41:17 +00:00
committed by GitHub
co-authored by Andrii Skliar
parent 7e583e615c
commit deb148e730
18 changed files with 1029 additions and 113 deletions
+4 -2
View File
@@ -153,6 +153,7 @@ def test_kda_recoverssm_derivation_is_revalidated():
use_replayssm=True,
use_kda_recoverssm=False,
mamba_cache_mode="none",
replayssm_buffer_len=16,
),
num_speculative_tokens=3,
model_config=SimpleNamespace(
@@ -184,8 +185,9 @@ def test_kda_recoverssm_derivation_is_revalidated():
config.cache_config.mamba_cache_mode = "none"
config.model_config.architecture = "NemotronHForCausalLM"
with pytest.raises(ValueError, match="only supported for Kimi-K3 KDA"):
VllmConfig.validate_mamba_cached_kernel(config)
config.mamba_config.backend = MambaBackendEnum.FLASHINFER
VllmConfig.validate_mamba_cached_kernel(config)
assert not config.cache_config.use_kda_recoverssm
config.model_config.architecture = "KimiLinearForCausalLM"
config.parallel_config.pipeline_parallel_size = 2