[Kernel][Model] Varlen prefill + Prefill chunking support for mamba kernels and Jamba model (#8533)

This commit is contained in:
Mor Zusman
2024-09-29 17:35:58 -04:00
committed by GitHub
parent 6c9ba48fde
commit f13a07b1f8
13 changed files with 1176 additions and 894 deletions
@@ -12,59 +12,44 @@ def causal_conv1d_fn(
x: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
seq_idx: Optional[torch.Tensor] = None,
initial_states: Optional[torch.Tensor] = None,
return_final_states: bool = False,
final_states_out=None,
activation: str = "silu",
query_start_loc: Optional[torch.Tensor] = None,
cache_indices: Optional[torch.Tensor] = None,
has_initial_state: Optional[torch.Tensor] = None,
conv_states: Optional[torch.Tensor] = None,
activation: Optional[str] = "silu",
):
"""
x: (batch, dim, seqlen)
x: (batch, dim, seqlen) or (dim,cu_seq_len) for varlen
sequences are concatenated from left to right for varlen
weight: (dim, width)
bias: (dim,)
seq_idx: (batch, seqlen)
initial_states: (batch, dim, width - 1)
final_states_out: (batch, dim, width - 1), to be written to
query_start_loc: (batch + 1) int32
The cumulative sequence lengths of the sequences in
the batch, used to index into sequence. prepended by 0.
for example: query_start_loc = torch.Tensor([0,10,16,17]),
x.shape=(dim,17)
cache_indices: (batch) int32
indicates the corresponding state index,
like so: conv_state = conv_states[cache_indices[batch_id]]
has_initial_state: (batch) bool
indicates whether should the kernel take the current state as initial
state for the calculations
conv_states: (...,dim,width - 1) itype
updated inplace if provided
activation: either None or "silu" or "swish"
out: (batch, dim, seqlen)
"""
if activation not in [None, "silu", "swish"]:
raise NotImplementedError("activation must be None, silu, or swish")
if x.stride(2) != 1 and x.stride(1) != 1:
if x.stride(-1) != 1:
x = x.contiguous()
bias = bias.contiguous() if bias is not None else None
if seq_idx is not None:
assert (initial_states is
None), "initial_states must be None if seq_idx is not None"
assert (not return_final_states
), "If seq_idx is not None, we don't return final_states_out"
seq_idx = seq_idx.contiguous() if seq_idx is not None else None
if initial_states is not None and (initial_states.stride(2) != 1
and initial_states.stride(1) != 1):
initial_states = initial_states.contiguous()
if return_final_states:
assert (
x.stride(1) == 1
), "Only channel-last layout support returning final_states_out"
if final_states_out is not None:
assert (final_states_out.stride(2) == 1
or final_states_out.stride(1) == 1)
else:
batch, dim, seqlen = x.shape
width = weight.shape[1]
final_states_out = torch.empty(batch,
width - 1,
dim,
device=x.device,
dtype=x.dtype).transpose(1, 2)
else:
final_states_out = None
out = ops.causal_conv1d_fwd(x, weight, bias, seq_idx, initial_states,
final_states_out, activation
out = ops.causal_conv1d_fwd(x, weight, bias, conv_states, query_start_loc,
cache_indices, has_initial_state, activation
in ["silu", "swish"])
return (out, None) if not return_final_states else (out, final_states_out)
return out
def causal_conv1d_update(x: torch.Tensor,
@@ -72,21 +57,33 @@ def causal_conv1d_update(x: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
activation: Optional[str] = None,
cache_seqlens: Optional[torch.Tensor] = None,
conv_state_indices: Optional[torch.Tensor] = None):
"""
x: (batch, dim)
conv_state: (batch, dim, width)
x: (batch, dim) or (batch, dim, seqlen)
conv_state: (batch, dim, state_len), where state_len >= width - 1
weight: (dim, width)
bias: (dim,)
cache_seqlens: (batch,), dtype int32.
If not None, the conv_state is treated as a circular buffer.
The conv_state will be updated by copying x to the conv_state
starting at the index
@cache_seqlens % state_len.
conv_state_indices: (batch,), dtype int32
If not None, the conv_state is a larger tensor along the batch dim,
and we are selecting the batch coords specified by conv_state_indices.
Useful for a continuous batching scenario.
out: (batch, dim)
out: (batch, dim) or (batch, dim, seqlen)
"""
if activation not in [None, "silu", "swish"]:
raise NotImplementedError("activation must be None, silu, or swish")
activation_bool = activation in ["silu", "swish"]
return ops.causal_conv1d_update(x, conv_state, weight, bias,
activation_bool, conv_state_indices)
activation_val = activation in ["silu", "swish"]
unsqueeze = x.dim() == 2
if unsqueeze:
x = x.unsqueeze(-1)
out = ops.causal_conv1d_update(x, conv_state, weight, bias, activation_val,
cache_seqlens, conv_state_indices)
if unsqueeze:
out = out.squeeze(-1)
return out
@@ -1,6 +1,8 @@
# Copyright (c) 2024, Tri Dao, Albert Gu.
# Adapted from https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/triton/selective_state_update.py
from typing import Tuple
import torch
import triton
import triton.language as tl
@@ -317,20 +319,50 @@ def selective_state_update(state,
return out
def selective_scan_fn(u,
delta,
A,
B,
C,
D=None,
z=None,
delta_bias=None,
delta_softplus=False,
return_last_state=False,
position_indices=None,
prev_state=None):
"""if return_last_state is True, returns (out, last_state)
last_state has shape (batch, dim, dstate).
def selective_scan_fn(
u,
ssm_states,
delta,
A,
B,
C,
D=None,
z=None,
delta_bias=None,
delta_softplus=False,
query_start_loc=None,
cache_indices=None,
has_initial_state=None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
u: (dim, total_length) for varlen or (batch, dim, seqlen)
delta: (dim, total_length) for varlen or (batch, dim, seqlen)
A: (dim, dstate)
B: (ngroups, dstate, total_length) for varlen or
(batch,ngroups,dstate,seqlen)
C: (ngroups, dstate, total_length) for varlen or
(batch,ngroups,dstate,seqlen)
D: (dim,)
z: (dim, total_length) for varlen or (batch, dim, seqlen)
dt_bias: (dim,) or (dim)
query_start_loc: (batch + 1) int32
The cumulative sequence lengths of the sequences in
the batch, used to index into sequence. prepended with 0.
for example: query_start_loc = torch.Tensor([0,10,16,17]),
x.shape=(dim,17)
cache_indices: (batch) int32
A tensor with each cell is a correspondent
input and output ssm_state index
has_initial_state: (batch) bool
A tensor populated with ones and zeros,
indicate if the ssm_state at the corresponding index should be
used as initial state. Not providing argument assumes
there's no initial state
returns
output: (dim, total_length) for varlen or (batch, dim, seqlen)
supports inplace replacement
last_state has shape (batch, dim, dstate).
supports inplace replacement if ssm_state was provided
"""
if u.stride(-1) != 1:
u = u.contiguous()
@@ -344,28 +376,20 @@ def selective_scan_fn(u,
C = C.contiguous()
if z is not None and z.stride(-1) != 1:
z = z.contiguous()
if B.dim() == 3:
if B.dim() == 3 and query_start_loc is None:
B = B.unsqueeze(1)
if C.dim() == 3:
if B.dim() == 2 and query_start_loc is not None:
B = B.unsqueeze(0)
if C.dim() == 3 and query_start_loc is None:
C = C.unsqueeze(1)
n_chunks = int((u.shape[-1] + 2048 - 1) / 2048)
x = torch.zeros((
u.shape[0],
u.shape[1],
n_chunks,
int(A.shape[1] * 2),
),
device=u.device,
dtype=torch.float32,
requires_grad=False)
x[:, :, 0, 0::2] = 1
if prev_state is not None:
x[:, :, 0, 1::2].copy_(prev_state)
out, *rest = ops.selective_scan_fwd(u, delta, A, B, C, D, z, delta_bias,
delta_softplus, position_indices, x)
last_state = x[:, :, -1, 1::2] # (batch, dim, dstate)
if C.dim() == 2 and query_start_loc is not None:
C = C.unsqueeze(0)
ops.selective_scan_fwd(u, delta, A, B, C, D, z, delta_bias, delta_softplus,
query_start_loc, cache_indices, has_initial_state,
ssm_states)
if z is None:
return out if not return_last_state else (out, last_state)
return delta # output written inplace to delta
else:
out_z = rest[0]
return out_z if not return_last_state else (out_z, last_state)
return z # output written inplace to z
+67 -97
View File
@@ -138,42 +138,47 @@ class JambaMambaMixer(nn.Module):
self.c_layernorm = RMSNorm(self.ssm_state_size,
eps=config.rms_norm_eps)
def mamba_forward(self,
hidden_states: torch.Tensor,
cache_params: MambaCacheParams = None):
def forward(self, hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata, conv_state: torch.Tensor,
ssm_state: torch.Tensor):
# 1. Gated MLP's linear projection
projected_states = self.in_proj(hidden_states)[0].transpose(1, 2)
hidden_states, gate = projected_states.chunk(2, dim=1)
projected_states = self.in_proj(hidden_states)[0].transpose(-2, -1)
hidden_states, gate = projected_states.chunk(2, dim=-2)
# 2. Convolution sequence transformation
conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0),
self.conv1d.weight.size(2))
if cache_params is not None and not cache_params.is_prompt:
hidden_states = causal_conv1d_update(
hidden_states.squeeze(-1),
cache_params.conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
)
hidden_states = hidden_states.unsqueeze(-1)
else:
if cache_params is not None:
conv_states = nn.functional.pad(
hidden_states,
(self.conv_kernel_size - hidden_states.shape[-1], 0))
cache_params.conv_state.copy_(conv_states)
hidden_states, _ = causal_conv1d_fn(
if attn_metadata.query_start_loc is not None \
and attn_metadata.context_lens_tensor is not None:
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ---------------------|
# |-- query_len ---|
hidden_states = causal_conv1d_fn(
hidden_states,
conv_weights,
self.conv1d.bias,
activation=self.activation,
conv_states=conv_state,
has_initial_state=attn_metadata.context_lens_tensor > 0,
query_start_loc=attn_metadata.query_start_loc)
else:
hidden_states = causal_conv1d_update(
hidden_states.transpose(0, 1),
conv_state,
conv_weights,
self.conv1d.bias,
self.activation,
)
hidden_states = hidden_states.transpose(0, 1)
# 3. State Space Model sequence transformation
# 3.a. input varying initialization of time_step, B and C
ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))[0]
ssm_parameters = self.x_proj(hidden_states.transpose(-2, -1))[0]
time_step, B, C = torch.split(
ssm_parameters,
@@ -184,72 +189,46 @@ class JambaMambaMixer(nn.Module):
B = self.b_layernorm(B.contiguous())
C = self.c_layernorm(C.contiguous())
discrete_time_step = self.dt_proj(time_step)[0].transpose(1, 2)
discrete_time_step = self.dt_proj(time_step)[0].transpose(-2, -1)
# 3.c perform the recurrence y ← SSM(A, B, C)(x)
time_proj_bias = (self.dt_proj.bias.float() if hasattr(
self.dt_proj, "bias") else None)
if cache_params is not None and not cache_params.is_prompt:
scan_outputs = selective_state_update(
cache_params.ssm_state,
hidden_states[..., 0],
discrete_time_step[..., 0],
self.A,
B[:, 0],
C[:, 0],
self.D,
gate[..., 0],
time_proj_bias,
dt_softplus=True,
).unsqueeze(-1)
else:
scan_outputs, ssm_state = selective_scan_fn(
if attn_metadata.query_start_loc is not None \
and attn_metadata.context_lens_tensor is not None:
scan_outputs = selective_scan_fn(
hidden_states,
ssm_state,
discrete_time_step,
self.A,
B.transpose(1, 2),
C.transpose(1, 2),
B.transpose(-2, -1),
C.transpose(-2, -1),
self.D.float(),
gate,
time_proj_bias,
delta_softplus=True,
return_last_state=True,
has_initial_state=attn_metadata.context_lens_tensor > 0,
query_start_loc=attn_metadata.query_start_loc)
else:
scan_outputs = selective_state_update(
ssm_state,
hidden_states.transpose(0, 1),
discrete_time_step.transpose(0, 1),
self.A,
B,
C,
self.D,
gate.transpose(0, 1),
time_proj_bias,
dt_softplus=True,
)
if ssm_state is not None and cache_params is not None:
cache_params.ssm_state.copy_(ssm_state)
scan_outputs = scan_outputs.transpose(0, 1)
# 4. Final linear projection
contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))[0]
contextualized_states = self.out_proj(scan_outputs.transpose(-2,
-1))[0]
return contextualized_states
def forward(
self,
hidden_states: torch.Tensor,
attn_metadata: AttentionMetadata,
conv_state: torch.Tensor,
ssm_state: torch.Tensor,
):
if attn_metadata.prefill_metadata is not None:
offset = 0
for i, prompt_len in enumerate(
attn_metadata.prefill_metadata.seq_lens):
cache = MambaCacheParams(True,
conv_state=conv_state[i].unsqueeze(0),
ssm_state=ssm_state[i].unsqueeze(0))
hidden_states[offset:offset + prompt_len].copy_(
self.mamba_forward(hidden_states[offset:offset +
prompt_len].unsqueeze(0),
cache_params=cache)[0])
offset += prompt_len
else:
cache = MambaCacheParams(False,
conv_state=conv_state,
ssm_state=ssm_state)
hidden_states = self.mamba_forward(hidden_states.unsqueeze(1),
cache_params=cache)
hidden_states = hidden_states.squeeze(1)
return hidden_states
class JambaMoE(nn.Module):
@@ -571,8 +550,6 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
lora_config: Optional[LoRAConfig] = None,
scheduler_config: Optional[SchedulerConfig] = None,
) -> None:
assert not scheduler_config.chunked_prefill_enabled, \
"Jamba currently does not support chunked prefill"
assert not cache_config.enable_prefix_caching, \
"Jamba currently does not support prefix caching"
@@ -616,18 +593,10 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
if "seqlen_agnostic_capture_inputs" not in kwargs:
# We get here only on Prefill/Eager mode runs
assert all(
key in kwargs
for key in ["request_ids_to_seq_ids", "finished_requests_ids"])
request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
finished_requests_ids = kwargs["finished_requests_ids"]
self._release_mamba_cache(finished_requests_ids)
batch_size = input_ids.shape[0]
if attn_metadata.prefill_metadata:
batch_size = len(request_ids_to_seq_ids)
mamba_cache = self._prepare_current_run_mamba_cache(
request_ids_to_seq_ids, batch_size, finished_requests_ids)
mamba_cache = self._release_finished_and_prepare_mamba_cache(
finished_requests_ids, request_ids_to_seq_ids)
else:
# CUDA graph capturing runs
mamba_cache = kwargs["seqlen_agnostic_capture_inputs"]
@@ -699,13 +668,15 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
def _prepare_current_run_mamba_cache(
self, request_ids_to_seq_ids: Dict[str, list[int]],
batch_size: int, finished_requests_ids: List[str]):
finished_requests_ids: List[str]
) -> Tuple[torch.Tensor, torch.Tensor]:
running_indices = []
request_ids_to_seq_ids_flatten = [
(req_id, seq_id)
for req_id, seq_ids in request_ids_to_seq_ids.items()
for seq_id in seq_ids
]
batch_size = len(request_ids_to_seq_ids_flatten)
for dest_index, (request_id,
seq_id) in enumerate(request_ids_to_seq_ids_flatten):
if request_id in finished_requests_ids:
@@ -769,22 +740,21 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
seq_ids2index.update({seq_id: to_index})
return
def _release_finished_and_prepare_mamba_cache(
self, finished_requests_ids,
request_ids_to_seq_ids) -> Tuple[torch.Tensor, torch.Tensor]:
self._release_mamba_cache(finished_requests_ids)
return self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
finished_requests_ids)
def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs):
"""
Copy the relevant Mamba cache into the CUDA graph input buffer
that was provided during the capture runs
(JambaForCausalLM.mamba_gc_cache_buffer).
"""
assert all(
key in kwargs
for key in ["request_ids_to_seq_ids", "finished_requests_ids"])
finished_requests_ids = kwargs["finished_requests_ids"]
self._release_mamba_cache(finished_requests_ids)
request_ids_to_seq_ids = kwargs["request_ids_to_seq_ids"]
cg_batch_size = input_buffers['input_ids'].shape[0]
self._prepare_current_run_mamba_cache(request_ids_to_seq_ids,
cg_batch_size,
finished_requests_ids)
self._release_finished_and_prepare_mamba_cache(
kwargs["finished_requests_ids"], kwargs["request_ids_to_seq_ids"])
def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
"""
@@ -819,7 +789,7 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA):
hidden_size = self.config.hidden_size
conv_state_shape = (
self.config.mamba_expand * hidden_size // world_size,
self.config.mamba_d_conv,
self.config.mamba_d_conv - 1,
)
temporal_state_shape = (
self.config.mamba_expand * self.config.hidden_size // world_size,