Skip to content

[Refactor] PagedKVCache spec for MLC-LLM #3203

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[Refactor] PagedKVCache spec for MLC-LLM #3203

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 7 additions & 7 deletions python/mlc_llm/model/baichuan/baichuan_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
from tvm.relax.frontend.nn import Tensor, op

from mlc_llm import op as op_ext
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.support import logging
from mlc_llm.support import tensor_parallel as tp
from mlc_llm.support.config import ConfigBase
Expand Down Expand Up @@ -280,7 +280,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mha",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -309,15 +309,15 @@ def get_default_spec(self):
},
"prefill": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode": {
"input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -326,23 +326,23 @@ def get_default_spec(self):
"batch_prefill": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode": {
"input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_verify": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
14 changes: 7 additions & 7 deletions python/mlc_llm/model/chatglm3/chatglm3_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
from tvm.relax.frontend.nn import Tensor, op

from mlc_llm import op as op_ext
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.support import logging
from mlc_llm.support import tensor_parallel as tp
from mlc_llm.support.config import ConfigBase
Expand Down Expand Up @@ -355,7 +355,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mha",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -384,15 +384,15 @@ def get_default_spec(self):
},
"prefill": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode": {
"input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -401,23 +401,23 @@ def get_default_spec(self):
"batch_prefill": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode": {
"input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_verify": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
14 changes: 7 additions & 7 deletions python/mlc_llm/model/cohere/cohere_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
from tvm.relax.frontend.nn import Tensor, op

from mlc_llm import op as op_ext
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.support import logging
from mlc_llm.support import tensor_parallel as tp
from mlc_llm.support.config import ConfigBase
Expand Down Expand Up @@ -324,7 +324,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mha",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -353,15 +353,15 @@ def get_default_spec(self):
},
"prefill": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode": {
"input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -370,23 +370,23 @@ def get_default_spec(self):
"batch_prefill": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode": {
"input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_verify": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
14 changes: 7 additions & 7 deletions python/mlc_llm/model/deepseek/deepseek_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@
from tvm.relax.frontend.nn import Tensor, op

from mlc_llm import op as op_ext
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.nn.expert import MixtralExperts
from mlc_llm.support import logging
from mlc_llm.support import tensor_parallel as tp
Expand Down Expand Up @@ -430,7 +430,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mha",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -459,15 +459,15 @@ def get_default_spec(self):
},
"prefill": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode": {
"input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -476,23 +476,23 @@ def get_default_spec(self):
"batch_prefill": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode": {
"input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_verify": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
18 changes: 9 additions & 9 deletions python/mlc_llm/model/deepseek_v2/deepseek_v2_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
from tvm.relax.frontend.nn.llm import position_embedding

from mlc_llm import op as op_ext
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.nn.expert import MixtralExperts
from mlc_llm.op import batch_matmul
from mlc_llm.support import logging
Expand Down Expand Up @@ -771,7 +771,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mla",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -802,23 +802,23 @@ def get_default_spec(self):
},
"prefill": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"extend": {
"input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode": {
"input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -827,7 +827,7 @@ def get_default_spec(self):
"batch_prefill": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand All @@ -836,23 +836,23 @@ def get_default_spec(self):
"batch_extend": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"logit_positions": nn.spec.Tensor(["batch_size"], "int32"),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode": {
"input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_verify": {
"input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
12 changes: 6 additions & 6 deletions python/mlc_llm/model/eagle/eagle_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@

from mlc_llm import op as op_ext
from mlc_llm.model.llama.llama_model import LlamaAttention, LlamaConfig, LlamaFFN
from mlc_llm.nn import PagedKVCache, RopeMode
from mlc_llm.nn import PagedKVCache, RopeMode, create_generic_paged_kv_cache
from mlc_llm.support import logging
from mlc_llm.support import tensor_parallel as tp

Expand Down Expand Up @@ -164,7 +164,7 @@ def create_paged_kv_cache( # pylint: disable=too-many-arguments
page_size: tir.Var,
support_sliding_window: tir.Var,
) -> PagedKVCache:
return PagedKVCache.create_generic(
return create_generic_paged_kv_cache(
attn_kind="mha",
max_batch_size=max_batch_size,
max_total_seq_len=max_total_seq_len,
Expand Down Expand Up @@ -201,31 +201,31 @@ def get_default_spec(self):
},
"prefill_to_last_hidden_states": {
"hidden_states": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"decode_to_last_hidden_states": {
"hidden_states": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_prefill_to_last_hidden_states": {
"hidden_states": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
},
},
"batch_decode_to_last_hidden_states": {
"hidden_states": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype),
"paged_kv_cache": nn.spec.Object(object_type=PagedKVCache),
"paged_kv_cache": nn.spec.PagedKVCache(),
"$": {
"param_mode": "packed",
"effect_mode": "none",
Expand Down
Loading
Loading