2023-11-11 06:04:50 +01:00
|
|
|
from __future__ import annotations
|
|
|
|
|
|
|
|
|
|
from typing import Sequence
|
|
|
|
|
|
|
|
|
|
from .constants import MODEL_ARCH, MODEL_TENSOR, MODEL_TENSORS, TENSOR_NAMES
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class TensorNameMap:
|
|
|
|
|
mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
|
|
|
|
|
# Token embeddings
|
|
|
|
|
MODEL_TENSOR.TOKEN_EMBD: (
|
|
|
|
|
"gpt_neox.embed_in", # gptneox
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx
|
2023-11-11 06:04:50 +01:00
|
|
|
"transformer.word_embeddings", # falcon
|
|
|
|
|
"word_embeddings", # bloom
|
|
|
|
|
"model.embed_tokens", # llama-hf
|
|
|
|
|
"tok_embeddings", # llama-pth
|
2024-02-13 18:03:53 +01:00
|
|
|
"embeddings.word_embeddings", # bert nomic-bert
|
2023-11-11 06:04:50 +01:00
|
|
|
"language_model.embedding.word_embeddings", # persimmon
|
2023-12-28 15:03:57 +01:00
|
|
|
"wte", # gpt2
|
2023-12-18 18:27:47 +01:00
|
|
|
"transformer.embd.wte", # phi2
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.tok_embeddings", # internlm2
|
llama : support Mamba Selective State Space Models (#5328)
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 23:31:00 +01:00
|
|
|
"model.embedding", # mamba-qbert
|
|
|
|
|
"backbone.embedding", # mamba
|
|
|
|
|
"backbone.embeddings", # mamba-hf
|
2024-03-23 17:41:53 +01:00
|
|
|
"transformer.in_out_embed", # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Token type embeddings
|
|
|
|
|
MODEL_TENSOR.TOKEN_TYPES: (
|
2024-02-13 18:03:53 +01:00
|
|
|
"embeddings.token_type_embeddings", # bert nomic-bert
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Normalization of token embeddings
|
|
|
|
|
MODEL_TENSOR.TOKEN_EMBD_NORM: (
|
|
|
|
|
"word_embeddings_layernorm", # bloom
|
2024-02-11 17:21:38 +01:00
|
|
|
"embeddings.LayerNorm", # bert
|
2024-02-13 18:03:53 +01:00
|
|
|
"emb_ln", # nomic-bert
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Position embeddings
|
|
|
|
|
MODEL_TENSOR.POS_EMBD: (
|
|
|
|
|
"transformer.wpe", # gpt2
|
|
|
|
|
"embeddings.position_embeddings", # bert
|
2023-12-28 15:03:57 +01:00
|
|
|
"wpe", # gpt2
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Output
|
|
|
|
|
MODEL_TENSOR.OUTPUT: (
|
|
|
|
|
"embed_out", # gptneox
|
2024-04-13 11:33:52 +02:00
|
|
|
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx
|
2024-02-01 10:19:51 +01:00
|
|
|
"output", # llama-pth bloom internlm2
|
2023-11-11 06:04:50 +01:00
|
|
|
"word_embeddings_for_head", # persimmon
|
2023-12-18 18:27:47 +01:00
|
|
|
"lm_head.linear", # phi2
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Output norm
|
|
|
|
|
MODEL_TENSOR.OUTPUT_NORM: (
|
|
|
|
|
"gpt_neox.final_layer_norm", # gptneox
|
|
|
|
|
"transformer.ln_f", # gpt2 gpt-j falcon
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.norm", # llama-hf baichuan internlm2
|
2023-11-11 06:04:50 +01:00
|
|
|
"norm", # llama-pth
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.norm_f", # mpt dbrx
|
2023-12-28 15:03:57 +01:00
|
|
|
"ln_f", # refact bloom qwen gpt2
|
2023-11-11 06:04:50 +01:00
|
|
|
"language_model.encoder.final_layernorm", # persimmon
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.final_layernorm", # persimmon
|
2023-12-18 18:27:47 +01:00
|
|
|
"lm_head.ln", # phi2
|
llama : support Mamba Selective State Space Models (#5328)
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 23:31:00 +01:00
|
|
|
"model.norm_f", # mamba-qbert
|
|
|
|
|
"backbone.norm_f", # mamba
|
2024-03-23 17:41:53 +01:00
|
|
|
"transformer.rms_norm", # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Rope frequencies
|
|
|
|
|
MODEL_TENSOR.ROPE_FREQS: (
|
|
|
|
|
"rope.freqs", # llama-pth
|
|
|
|
|
),
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
|
|
|
|
|
# Attention norm
|
|
|
|
|
MODEL_TENSOR.ATTN_NORM: (
|
|
|
|
|
"gpt_neox.layers.{bid}.input_layernorm", # gptneox
|
2023-12-01 19:16:31 +01:00
|
|
|
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen
|
2023-11-11 06:04:50 +01:00
|
|
|
"transformer.blocks.{bid}.norm_1", # mpt
|
|
|
|
|
"transformer.h.{bid}.input_layernorm", # falcon7b
|
|
|
|
|
"h.{bid}.input_layernorm", # bloom
|
|
|
|
|
"transformer.h.{bid}.ln_mlp", # falcon40b
|
|
|
|
|
"model.layers.{bid}.input_layernorm", # llama-hf
|
|
|
|
|
"layers.{bid}.attention_norm", # llama-pth
|
|
|
|
|
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
|
|
|
|
|
"model.layers.{bid}.ln1", # yi
|
2023-12-28 15:03:57 +01:00
|
|
|
"h.{bid}.ln_1", # gpt2
|
2023-12-18 18:27:47 +01:00
|
|
|
"transformer.h.{bid}.ln", # phi2
|
2023-12-24 14:35:49 +01:00
|
|
|
"model.layers.layers.{bid}.norm", # plamo
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.layers.{bid}.attention_norm", # internlm2
|
llama : support Mamba Selective State Space Models (#5328)
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 23:31:00 +01:00
|
|
|
"model.layers.{bid}.norm", # mamba-qbert
|
|
|
|
|
"backbone.layers.{bid}.norm", # mamba
|
2024-03-23 17:41:53 +01:00
|
|
|
"transformer.decoder_layer.{bid}.rms_norm", # Grok
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention norm 2
|
|
|
|
|
MODEL_TENSOR.ATTN_NORM_2: (
|
|
|
|
|
"transformer.h.{bid}.ln_attn", # falcon40b
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention query-key-value
|
|
|
|
|
MODEL_TENSOR.ATTN_QKV: (
|
|
|
|
|
"gpt_neox.layers.{bid}.attention.query_key_value", # gptneox
|
2023-12-01 19:16:31 +01:00
|
|
|
"transformer.h.{bid}.attn.c_attn", # gpt2 qwen
|
2023-11-11 06:04:50 +01:00
|
|
|
"transformer.blocks.{bid}.attn.Wqkv", # mpt
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx
|
2023-11-11 06:04:50 +01:00
|
|
|
"transformer.h.{bid}.self_attention.query_key_value", # falcon
|
|
|
|
|
"h.{bid}.self_attention.query_key_value", # bloom
|
|
|
|
|
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.layers.{bid}.self_attn.query_key_value", # persimmon
|
2023-12-28 15:03:57 +01:00
|
|
|
"h.{bid}.attn.c_attn", # gpt2
|
2023-12-18 18:27:47 +01:00
|
|
|
"transformer.h.{bid}.mixer.Wqkv", # phi2
|
2024-02-13 18:03:53 +01:00
|
|
|
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention query
|
|
|
|
|
MODEL_TENSOR.ATTN_Q: (
|
2024-03-23 17:41:53 +01:00
|
|
|
"model.layers.{bid}.self_attn.q_proj", # llama-hf
|
|
|
|
|
"layers.{bid}.attention.wq", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.attention.self.query", # bert
|
|
|
|
|
"transformer.h.{bid}.attn.q_proj", # gpt-j
|
|
|
|
|
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
|
|
|
|
|
"model.layers.{bid}.attention.wq", # internlm2
|
|
|
|
|
"transformer.decoder_layer.{bid}.multi_head_attention.query" # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention key
|
|
|
|
|
MODEL_TENSOR.ATTN_K: (
|
2024-03-23 17:41:53 +01:00
|
|
|
"model.layers.{bid}.self_attn.k_proj", # llama-hf
|
|
|
|
|
"layers.{bid}.attention.wk", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.attention.self.key", # bert
|
|
|
|
|
"transformer.h.{bid}.attn.k_proj", # gpt-j
|
|
|
|
|
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
|
|
|
|
|
"model.layers.{bid}.attention.wk", # internlm2
|
|
|
|
|
"transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention value
|
|
|
|
|
MODEL_TENSOR.ATTN_V: (
|
2024-03-23 17:41:53 +01:00
|
|
|
"model.layers.{bid}.self_attn.v_proj", # llama-hf
|
|
|
|
|
"layers.{bid}.attention.wv", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.attention.self.value", # bert
|
|
|
|
|
"transformer.h.{bid}.attn.v_proj", # gpt-j
|
|
|
|
|
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
|
|
|
|
|
"model.layers.{bid}.attention.wv", # internlm2
|
|
|
|
|
"transformer.decoder_layer.{bid}.multi_head_attention.value" # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Attention output
|
|
|
|
|
MODEL_TENSOR.ATTN_OUT: (
|
2024-04-13 11:33:52 +02:00
|
|
|
"gpt_neox.layers.{bid}.attention.dense", # gptneox
|
|
|
|
|
"transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen
|
|
|
|
|
"transformer.blocks.{bid}.attn.out_proj", # mpt
|
|
|
|
|
"transformer.h.{bid}.self_attention.dense", # falcon
|
|
|
|
|
"h.{bid}.self_attention.dense", # bloom
|
|
|
|
|
"model.layers.{bid}.self_attn.o_proj", # llama-hf
|
|
|
|
|
"layers.{bid}.attention.wo", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.attention.output.dense", # bert
|
|
|
|
|
"transformer.h.{bid}.attn.out_proj", # gpt-j
|
|
|
|
|
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
|
|
|
|
|
"model.layers.{bid}.self_attn.dense", # persimmon
|
|
|
|
|
"h.{bid}.attn.c_proj", # gpt2
|
|
|
|
|
"transformer.h.{bid}.mixer.out_proj", # phi2
|
|
|
|
|
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
|
|
|
|
|
"model.layers.{bid}.attention.wo", # internlm2
|
|
|
|
|
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
|
|
|
|
|
"transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok
|
|
|
|
|
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
2024-02-11 17:21:38 +01:00
|
|
|
# Attention output norm
|
|
|
|
|
MODEL_TENSOR.ATTN_OUT_NORM: (
|
|
|
|
|
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
|
2024-02-13 18:03:53 +01:00
|
|
|
"encoder.layers.{bid}.norm1", # nomic-bert
|
2024-03-23 17:41:53 +01:00
|
|
|
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx
|
2024-02-11 17:21:38 +01:00
|
|
|
),
|
|
|
|
|
|
2023-11-11 06:04:50 +01:00
|
|
|
# Rotary embeddings
|
|
|
|
|
MODEL_TENSOR.ATTN_ROT_EMBD: (
|
2023-12-24 14:35:49 +01:00
|
|
|
"model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf
|
|
|
|
|
"layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
|
|
|
|
|
"model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo
|
2024-01-19 10:07:27 +01:00
|
|
|
"transformer.h.{bid}.attn.rotary_emb.inv_freq", # codeshell
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Feed-forward norm
|
|
|
|
|
MODEL_TENSOR.FFN_NORM: (
|
|
|
|
|
"gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox
|
2023-12-01 19:16:31 +01:00
|
|
|
"transformer.h.{bid}.ln_2", # gpt2 refact qwen
|
2023-11-11 06:04:50 +01:00
|
|
|
"h.{bid}.post_attention_layernorm", # bloom
|
|
|
|
|
"transformer.blocks.{bid}.norm_2", # mpt
|
|
|
|
|
"model.layers.{bid}.post_attention_layernorm", # llama-hf
|
|
|
|
|
"layers.{bid}.ffn_norm", # llama-pth
|
|
|
|
|
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
|
|
|
|
|
"model.layers.{bid}.ln2", # yi
|
2023-12-28 15:03:57 +01:00
|
|
|
"h.{bid}.ln_2", # gpt2
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.layers.{bid}.ffn_norm", # internlm2
|
2024-03-23 17:41:53 +01:00
|
|
|
"transformer.decoder_layer.{bid}.rms_norm_2", # Grok
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
2023-12-13 13:04:25 +01:00
|
|
|
MODEL_TENSOR.FFN_GATE_INP: (
|
2024-04-13 11:33:52 +02:00
|
|
|
"layers.{bid}.feed_forward.gate", # mixtral
|
|
|
|
|
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
|
2024-04-16 17:40:48 +02:00
|
|
|
"model.layers.{bid}.mlp.gate", # qwen2moe
|
2024-04-13 11:33:52 +02:00
|
|
|
"transformer.decoder_layer.{bid}.router", # Grok
|
|
|
|
|
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
|
2023-12-13 13:04:25 +01:00
|
|
|
),
|
|
|
|
|
|
2024-04-16 17:40:48 +02:00
|
|
|
MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
|
|
|
|
|
"model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
|
|
|
|
|
),
|
|
|
|
|
|
2023-11-11 06:04:50 +01:00
|
|
|
# Feed-forward up
|
|
|
|
|
MODEL_TENSOR.FFN_UP: (
|
|
|
|
|
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
|
|
|
|
|
"transformer.h.{bid}.mlp.c_fc", # gpt2
|
|
|
|
|
"transformer.blocks.{bid}.ffn.up_proj", # mpt
|
|
|
|
|
"transformer.h.{bid}.mlp.dense_h_to_4h", # falcon
|
|
|
|
|
"h.{bid}.mlp.dense_h_to_4h", # bloom
|
|
|
|
|
"model.layers.{bid}.mlp.up_proj", # llama-hf refact
|
|
|
|
|
"layers.{bid}.feed_forward.w3", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.intermediate.dense", # bert
|
|
|
|
|
"transformer.h.{bid}.mlp.fc_in", # gpt-j
|
|
|
|
|
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.layers.{bid}.mlp.dense_h_to_4h", # persimmon
|
2023-12-01 19:16:31 +01:00
|
|
|
"transformer.h.{bid}.mlp.w1", # qwen
|
2023-12-28 15:03:57 +01:00
|
|
|
"h.{bid}.mlp.c_fc", # gpt2
|
2023-12-18 18:27:47 +01:00
|
|
|
"transformer.h.{bid}.mlp.fc1", # phi2
|
2024-01-13 12:44:37 +01:00
|
|
|
"model.layers.{bid}.mlp.fc1", # phi2
|
2023-12-24 14:35:49 +01:00
|
|
|
"model.layers.layers.{bid}.mlp.up_proj", # plamo
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.layers.{bid}.feed_forward.w3", # internlm2
|
2024-02-13 18:03:53 +01:00
|
|
|
"encoder.layers.{bid}.mlp.fc11", # nomic-bert
|
2024-03-01 20:30:46 +01:00
|
|
|
"model.layers.{bid}.mlp.c_fc", # starcoder2
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
2023-12-13 13:04:25 +01:00
|
|
|
MODEL_TENSOR.FFN_UP_EXP: (
|
2024-04-16 17:40:48 +02:00
|
|
|
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
|
|
|
|
|
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
|
|
|
|
|
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
|
|
|
|
|
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe (merged)
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.FFN_UP_SHEXP: (
|
|
|
|
|
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
|
2023-12-13 13:04:25 +01:00
|
|
|
),
|
|
|
|
|
|
2023-12-27 16:39:45 +01:00
|
|
|
# AWQ-activation gate
|
|
|
|
|
MODEL_TENSOR.FFN_ACT: (
|
|
|
|
|
"transformer.blocks.{bid}.ffn.act", # mpt
|
|
|
|
|
),
|
|
|
|
|
|
2023-11-11 06:04:50 +01:00
|
|
|
# Feed-forward gate
|
|
|
|
|
MODEL_TENSOR.FFN_GATE: (
|
2023-12-13 13:04:25 +01:00
|
|
|
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
|
|
|
|
|
"layers.{bid}.feed_forward.w1", # llama-pth
|
|
|
|
|
"transformer.h.{bid}.mlp.w2", # qwen
|
2023-12-24 14:35:49 +01:00
|
|
|
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.layers.{bid}.feed_forward.w1", # internlm2
|
2024-02-13 18:03:53 +01:00
|
|
|
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
|
2023-12-13 13:04:25 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.FFN_GATE_EXP: (
|
2024-04-13 11:33:52 +02:00
|
|
|
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
|
|
|
|
|
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
|
|
|
|
|
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
|
2024-04-16 17:40:48 +02:00
|
|
|
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe (merged)
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.FFN_GATE_SHEXP: (
|
|
|
|
|
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
# Feed-forward down
|
|
|
|
|
MODEL_TENSOR.FFN_DOWN: (
|
|
|
|
|
"gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox
|
2023-12-01 19:16:31 +01:00
|
|
|
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen
|
2023-11-11 06:04:50 +01:00
|
|
|
"transformer.blocks.{bid}.ffn.down_proj", # mpt
|
|
|
|
|
"transformer.h.{bid}.mlp.dense_4h_to_h", # falcon
|
|
|
|
|
"h.{bid}.mlp.dense_4h_to_h", # bloom
|
|
|
|
|
"model.layers.{bid}.mlp.down_proj", # llama-hf
|
|
|
|
|
"layers.{bid}.feed_forward.w2", # llama-pth
|
|
|
|
|
"encoder.layer.{bid}.output.dense", # bert
|
|
|
|
|
"transformer.h.{bid}.mlp.fc_out", # gpt-j
|
|
|
|
|
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.layers.{bid}.mlp.dense_4h_to_h", # persimmon
|
2023-12-28 15:03:57 +01:00
|
|
|
"h.{bid}.mlp.c_proj", # gpt2
|
2023-12-18 18:27:47 +01:00
|
|
|
"transformer.h.{bid}.mlp.fc2", # phi2
|
2024-01-13 12:44:37 +01:00
|
|
|
"model.layers.{bid}.mlp.fc2", # phi2
|
2023-12-24 14:35:49 +01:00
|
|
|
"model.layers.layers.{bid}.mlp.down_proj", # plamo
|
2024-02-01 10:19:51 +01:00
|
|
|
"model.layers.{bid}.feed_forward.w2", # internlm2
|
2024-02-13 18:03:53 +01:00
|
|
|
"encoder.layers.{bid}.mlp.fc2", # nomic-bert
|
2024-03-01 20:30:46 +01:00
|
|
|
"model.layers.{bid}.mlp.c_proj", # starcoder2
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
2023-12-13 13:04:25 +01:00
|
|
|
MODEL_TENSOR.FFN_DOWN_EXP: (
|
2024-04-16 17:40:48 +02:00
|
|
|
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
|
|
|
|
|
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
|
|
|
|
|
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
|
|
|
|
|
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe (merged)
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.FFN_DOWN_SHEXP: (
|
|
|
|
|
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
|
2023-12-13 13:04:25 +01:00
|
|
|
),
|
|
|
|
|
|
2023-11-11 06:04:50 +01:00
|
|
|
MODEL_TENSOR.ATTN_Q_NORM: (
|
|
|
|
|
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
|
2024-04-09 10:16:13 +02:00
|
|
|
"model.layers.{bid}.self_attn.q_norm", # cohere
|
2024-04-03 20:05:10 +02:00
|
|
|
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.ATTN_K_NORM: (
|
|
|
|
|
"language_model.encoder.layers.{bid}.self_attention.k_layernorm",
|
2024-01-12 12:01:56 +01:00
|
|
|
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
|
2024-04-09 10:16:13 +02:00
|
|
|
"model.layers.{bid}.self_attn.k_norm", # cohere
|
2024-04-03 20:05:10 +02:00
|
|
|
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
|
2023-11-11 06:04:50 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.ROPE_FREQS: (
|
|
|
|
|
"language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon
|
|
|
|
|
),
|
2024-02-11 17:21:38 +01:00
|
|
|
|
|
|
|
|
MODEL_TENSOR.LAYER_OUT_NORM: (
|
2024-03-23 17:41:53 +01:00
|
|
|
"encoder.layer.{bid}.output.LayerNorm", # bert
|
|
|
|
|
"encoder.layers.{bid}.norm2", # nomic-bert
|
|
|
|
|
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
|
llama : support Mamba Selective State Space Models (#5328)
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 23:31:00 +01:00
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_IN: (
|
|
|
|
|
"model.layers.{bid}.in_proj",
|
|
|
|
|
"backbone.layers.{bid}.mixer.in_proj",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_CONV1D: (
|
|
|
|
|
"model.layers.{bid}.conv1d",
|
|
|
|
|
"backbone.layers.{bid}.mixer.conv1d",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_X: (
|
|
|
|
|
"model.layers.{bid}.x_proj",
|
|
|
|
|
"backbone.layers.{bid}.mixer.x_proj",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_DT: (
|
|
|
|
|
"model.layers.{bid}.dt_proj",
|
|
|
|
|
"backbone.layers.{bid}.mixer.dt_proj",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_A: (
|
|
|
|
|
"model.layers.{bid}.A_log",
|
|
|
|
|
"backbone.layers.{bid}.mixer.A_log",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_D: (
|
|
|
|
|
"model.layers.{bid}.D",
|
|
|
|
|
"backbone.layers.{bid}.mixer.D",
|
|
|
|
|
),
|
|
|
|
|
|
|
|
|
|
MODEL_TENSOR.SSM_OUT: (
|
|
|
|
|
"model.layers.{bid}.out_proj",
|
|
|
|
|
"backbone.layers.{bid}.mixer.out_proj",
|
|
|
|
|
),
|
2023-11-11 06:04:50 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mapping: dict[str, tuple[MODEL_TENSOR, str]]
|
|
|
|
|
|
|
|
|
|
def __init__(self, arch: MODEL_ARCH, n_blocks: int):
|
|
|
|
|
self.mapping = {}
|
|
|
|
|
for tensor, keys in self.mappings_cfg.items():
|
|
|
|
|
if tensor not in MODEL_TENSORS[arch]:
|
|
|
|
|
continue
|
|
|
|
|
tensor_name = TENSOR_NAMES[tensor]
|
|
|
|
|
self.mapping[tensor_name] = (tensor, tensor_name)
|
|
|
|
|
for key in keys:
|
|
|
|
|
self.mapping[key] = (tensor, tensor_name)
|
|
|
|
|
for bid in range(n_blocks):
|
|
|
|
|
for tensor, keys in self.block_mappings_cfg.items():
|
|
|
|
|
if tensor not in MODEL_TENSORS[arch]:
|
|
|
|
|
continue
|
2023-12-13 13:04:25 +01:00
|
|
|
# TODO: make this configurable
|
2024-04-16 17:40:48 +02:00
|
|
|
n_experts = 60
|
2023-12-13 13:04:25 +01:00
|
|
|
for xid in range(n_experts):
|
|
|
|
|
tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
|
|
|
|
|
self.mapping[tensor_name] = (tensor, tensor_name)
|
|
|
|
|
for key in keys:
|
|
|
|
|
key = key.format(bid = bid, xid = xid)
|
|
|
|
|
self.mapping[key] = (tensor, tensor_name)
|
2023-11-11 06:04:50 +01:00
|
|
|
|
|
|
|
|
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
|
|
|
|
|
result = self.mapping.get(key)
|
|
|
|
|
if result is not None:
|
|
|
|
|
return result
|
|
|
|
|
for suffix in try_suffixes:
|
|
|
|
|
if key.endswith(suffix):
|
|
|
|
|
result = self.mapping.get(key[:-len(suffix)])
|
|
|
|
|
if result is not None:
|
|
|
|
|
return result[0], result[1] + suffix
|
|
|
|
|
return None
|
|
|
|
|
|
|
|
|
|
def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
|
|
|
|
|
result = self.get_type_and_name(key, try_suffixes = try_suffixes)
|
|
|
|
|
if result is None:
|
|
|
|
|
return None
|
|
|
|
|
return result[1]
|
|
|
|
|
|
|
|
|
|
def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
|
|
|
|
|
result = self.get_type_and_name(key, try_suffixes = try_suffixes)
|
|
|
|
|
if result is None:
|
|
|
|
|
return None
|
|
|
|
|
return result[0]
|
|
|
|
|
|
|
|
|
|
def __getitem__(self, key: str) -> str:
|
|
|
|
|
try:
|
|
|
|
|
return self.mapping[key][1]
|
|
|
|
|
except KeyError:
|
|
|
|
|
raise KeyError(key)
|
|
|
|
|
|
|
|
|
|
def __contains__(self, key: str) -> bool:
|
|
|
|
|
return key in self.mapping
|
|
|
|
|
|
|
|
|
|
def __repr__(self) -> str:
|
|
|
|
|
return repr(self.mapping)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap:
|
|
|
|
|
return TensorNameMap(arch, n_blocks)
|