biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V1] [Hybrid] Support using float32 for state in Hybrid Models (Mamba2, Mamba1, Minimax) (#22928)

Browse Source 
Signed-off-by: Daniel Afrimi <danielafrimi8@gmail.com>
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Daniel Afrimi <danielafrimi8@gmail.com>
Co-authored-by: Burkhard Ringlein <ngl@zurich.ibm.com>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Thomas Parnell
2025-08-15 14:57:06 +02:00
committed by GitHub
parent 22341b996e
commit 75531a6c13
23 changed files with 467 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
vllm/model_executor/models/jamba.py
View File

@@ -10,7 +10,7 @@ from transformers import JambaConfig
from vllm import envs
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, VllmConfig
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -21,7 +21,7 @@ from vllm.model_executor.layers.linear import (QKVParallelLinear,
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.mamba_mixer import MambaMixer
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateShapeCalculator)
MambaStateDtypeCalculator, MambaStateShapeCalculator)
from vllm.model_executor.layers.pooler import DispatchPooler, Pooler
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -94,6 +94,7 @@ class JambaMambaDecoderLayer(nn.Module):
def __init__(self,
config: JambaConfig,
layer_idx: int,
model_config: Optional[ModelConfig] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
is_lora_enabled: Optional[bool] = False,
@@ -114,6 +115,8 @@ class JambaMambaDecoderLayer(nn.Module):
rms_norm_eps=config.rms_norm_eps,
activation=config.hidden_act,
is_lora_enabled = self.is_lora_enabled,
model_config=model_config,
cache_config=cache_config,
prefix=f"{prefix}.mixer",
)
@@ -164,6 +167,7 @@ class JambaAttentionDecoderLayer(nn.Module):
def __init__(self,
config: JambaConfig,
layer_idx: int,
model_config: Optional[ModelConfig] = None,
cache_config: Optional[CacheConfig] = None,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
@@ -280,6 +284,7 @@ class JambaModel(nn.Module):
super().__init__()
config = vllm_config.model_config.hf_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
@@ -304,6 +309,7 @@ class JambaModel(nn.Module):
config.layers_block_type[layer_idx]]
return layer_class(config,
layer_idx,
model_config,
cache_config,
quant_config=quant_config,
prefix=prefix,
@@ -520,9 +526,11 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
self.vllm_config.parallel_config, LayerBlockType.mamba)
state_shape = self.get_mamba_state_shape_from_config(
self.vllm_config)
state_dtype = self.get_mamba_state_dtype_from_config(
self.vllm_config)
self.mamba_cache = MambaCacheManager(self.vllm_config,
self.lm_head.weight.dtype,
num_layers, *state_shape)
num_layers, *state_shape,
*state_dtype)
mamba_cache_params = self.mamba_cache.current_run_tensors(**kwargs)
@@ -537,6 +545,18 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size)
@classmethod
def get_mamba_state_dtype_from_config(
cls,
vllm_config: "VllmConfig",
) -> tuple[torch.dtype, torch.dtype]:
return MambaStateDtypeCalculator.mamba1_state_dtype(
vllm_config.model_config.dtype,
vllm_config.cache_config.mamba_cache_dtype,
vllm_config.cache_config.mamba_ssm_cache_dtype,
)
@classmethod
def get_mamba_state_shape_from_config(
cls,