root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 02:14:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

llama : support Mamba Selective State Space Models (#5328)

Browse Source 
* 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`.
This commit is contained in:
compilade 2024-03-08 17:31:00 -05:00 committed by GitHub
parent 515f7d0d4f
commit c2101a2e90
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
15 changed files with 1342 additions and 74 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@ -10,6 +10,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
### Recent API changes
- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
@ -110,6 +111,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
- [x] [Gemma](https://ai.google.dev/gemma)
- [x] [Mamba](https://github.com/state-spaces/mamba)
**Multimodal models:**

1
common/common.cpp
View File

@ -1288,6 +1288,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.n_ctx = params.n_ctx;
cparams.n_batch = params.n_batch;
cparams.n_parallel = params.n_parallel;
cparams.n_threads = params.n_threads;
cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
cparams.seed = params.seed;

118
convert-hf-to-gguf.py
View File

@ -1847,6 +1847,124 @@ class StarCoder2Model(Model):
model_arch = gguf.MODEL_ARCH.STARCODER2
@Model.register("MambaForCausalLM", "MambaLMHeadModel")
class MambaModel(Model):
model_arch = gguf.MODEL_ARCH.MAMBA
def set_vocab(self):
vocab_size = self.hparams["vocab_size"]
# Round vocab size to next multiple of 8
pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
# pad using ceiling division
# ref: https://stackoverflow.com/a/17511341/22827863
vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
self.hparams["vocab_size"] = vocab_size
if (self.dir_model / "tokenizer.json").is_file():
self._set_vocab_gpt2()
else:
# Use the GPT-NeoX tokenizer when no tokenizer files are present
tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
neox_reader = gguf.GGUFReader(tokenizer_path, "r")
field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
def set_gguf_parameters(self):
d_model = self.find_hparam(["hidden_size", "d_model"])
d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
# ceiling division
# ref: https://stackoverflow.com/a/17511341/22827863
# ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
# Fail early for models which don't have a block expansion factor of 2
assert d_inner == 2 * d_model
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
self.gguf_writer.add_embedding_length(d_model)
self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
self.gguf_writer.add_block_count(self.hparams["n_layer"])
self.gguf_writer.add_ssm_conv_kernel(d_conv)
self.gguf_writer.add_ssm_inner_size(d_inner)
self.gguf_writer.add_ssm_state_size(d_state)
self.gguf_writer.add_ssm_time_step_rank(dt_rank)
self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams["n_layer"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
tok_embd = None
tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
output_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT] + ".weight"
for name, data_torch in self.get_tensors():
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
if name.endswith(".A_log"):
print("A_log --> A ==> " + new_name)
data_torch = -torch.exp(data_torch)
# assuming token_embd.weight is seen before output.weight
if tok_embd is not None and new_name == output_name:
if torch.equal(tok_embd, data_torch):
print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
continue
if new_name == tok_embd_name:
tok_embd = data_torch
data = data_torch.squeeze().numpy()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert big float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
###### CONVERSION LOGIC ######

13
examples/batched-bench/batched-bench.cpp
View File

@ -105,6 +105,9 @@ int main(int argc, char ** argv) {
ctx_params.n_threads = params.n_threads;
ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
// ensure enough sequences are available
ctx_params.n_parallel = *std::max_element(n_pl.begin(), n_pl.end());
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
if (ctx == NULL) {
@ -174,10 +177,10 @@ int main(int argc, char ** argv) {
llama_batch_clear(batch);
const int n_tokens = is_pp_shared ? pp : pl*pp;
for (int i = 0; i < n_tokens; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
for (int i = 0; i < pp; ++i) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
llama_batch_add(batch, 0, i, { j }, false);
}
}
batch.logits[batch.n_tokens - 1] = true;
@ -192,7 +195,7 @@ int main(int argc, char ** argv) {
if (is_pp_shared) {
for (int32_t i = 1; i < pl; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
}

3
examples/batched/batched.cpp
View File

@ -80,6 +80,7 @@ int main(int argc, char ** argv) {
ctx_params.seed = 1234;
ctx_params.n_ctx = n_kv_req;
ctx_params.n_batch = std::max(n_len, n_parallel);
ctx_params.n_parallel = n_parallel;
ctx_params.n_threads = params.n_threads;
ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
@ -132,7 +133,7 @@ int main(int argc, char ** argv) {
// assign the system KV cache to all parallel sequences
// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
for (int32_t i = 1; i < n_parallel; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
if (n_parallel > 1) {

20
examples/parallel/parallel.cpp
View File

@ -107,6 +107,9 @@ int main(int argc, char ** argv) {
// number of simultaneous "clients" to simulate
const int32_t n_clients = params.n_parallel;
// dedicate one sequence to the system prompt
params.n_parallel += 1;
// requests to simulate
const int32_t n_seq = params.n_sequences;
@ -196,8 +199,8 @@ int main(int argc, char ** argv) {
}
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < n_clients; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
for (int32_t i = 1; i <= n_clients; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("\n");
@ -221,15 +224,17 @@ int main(int argc, char ** argv) {
client.i_batch = batch.n_tokens;
llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
client.n_decoded += 1;
}
if (batch.n_tokens == 0) {
// all sequences have ended - clear the entire KV cache
for (int i = 0; i < n_clients; ++i) {
llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
for (int i = 1; i <= n_clients; ++i) {
llama_kv_cache_seq_rm(ctx, i, -1, -1);
// but keep the system prompt
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("%s: clearing the KV cache\n", __func__);
@ -255,7 +260,7 @@ int main(int argc, char ** argv) {
tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
for (size_t i = 0; i < tokens_prompt.size(); ++i) {
llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
}
// extract the logits only for the last token
@ -366,7 +371,8 @@ int main(int argc, char ** argv) {
}
// delete only the generated part of the sequence, i.e. keep the system prompt in the cache
llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
const auto t_main_end = ggml_time_us();

9
examples/perplexity/perplexity.cpp
View File

@ -809,7 +809,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
const int n_batch = params.n_batch;
const int max_tasks_per_batch = 32;
const int max_seq = 4*max_tasks_per_batch;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
@ -1086,7 +1086,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
const int n_batch = params.n_batch;
const int max_tasks_per_batch = 128;
const int max_seq = 2*max_tasks_per_batch;
const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
@ -1438,7 +1438,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
const int n_batch = params.n_batch;
const int max_tasks_per_batch = 32;
const int max_seq = 4*max_tasks_per_batch;
const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
@ -1815,6 +1815,9 @@ int main(int argc, char ** argv) {
llama_model * model;
llama_context * ctx;
// ensure there's at least enough seq_ids for HellaSwag
params.n_parallel = std::max(4, params.n_parallel);
// load the model and apply lora adapter, if any
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (model == NULL) {

51
examples/server/server.cpp
View File

@ -659,7 +659,11 @@ struct server_context {
bool load_model(const gpt_params & params_) {
params = params_;
// dedicate one sequence to the system prompt
params.n_parallel += 1;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
params.n_parallel -= 1; // but be sneaky about it
if (model == nullptr) {
LOG_ERROR("unable to load model", {{"model", params.model}});
return false;
@ -1018,8 +1022,8 @@ struct server_context {
}
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < params.n_parallel; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
for (int32_t i = 1; i <= params.n_parallel; ++i) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
}
@ -1306,7 +1310,7 @@ struct server_context {
const int n_embd = llama_n_embd(model);
for (int i = 0; i < batch.n_tokens; ++i) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) {
continue;
}
@ -1633,8 +1637,8 @@ struct server_context {
{"n_cache_tokens", slot.cache_tokens.size()}
});
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
llama_kv_cache_seq_rm (ctx, slot.id + 1, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id + 1, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
if (slot.params.cache_prompt) {
for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@ -1666,7 +1670,7 @@ struct server_context {
// TODO: we always have to take into account the "system_tokens"
// this is not great and needs to be improved somehow
llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true);
llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id + 1 }, true);
slot.n_past += 1;
@ -1804,9 +1808,6 @@ struct server_context {
// reuse any previously computed tokens that are common with the new prompt
slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
// remove the non-common part from the cache
slot.cache_tokens.resize(slot.n_past);
// push the prompt into the sampling context (do not apply grammar)
for (int i = 0; i < slot.n_past; ++i) {
llama_sampling_accept(slot.ctx_sampling, ctx, slot.cache_tokens[i], false);
@ -1837,8 +1838,28 @@ struct server_context {
}
}
const int p0 = (int) system_tokens.size() + slot.n_past;
llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
// keep only the common part
int p0 = (int) system_tokens.size() + slot.n_past;
if (!llama_kv_cache_seq_rm(ctx, slot.id + 1, p0, -1)) {
// could not partially delete (likely using a non-Transformer model)
llama_kv_cache_seq_rm(ctx, slot.id + 1, -1, -1);
p0 = (int) system_tokens.size();
if (p0 != 0) {
// copy over the system prompt when there is one
llama_kv_cache_seq_cp(ctx, 0, slot.id + 1, -1, -1);
}
// there is no common part left (except for the system prompt)
slot.n_past = 0;
slot.n_past_se = 0;
slot.ga_i = 0;
// TODO: is the system prompt ever in the sampling context?
llama_sampling_reset(slot.ctx_sampling);
}
// remove the non-common part from the cache
slot.cache_tokens.resize(slot.n_past);
LOG_INFO("kv cache rm [p0, end)", {
{ "id_slot", slot.id },
@ -1863,7 +1884,7 @@ struct server_context {
}
}
llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id }, false);
llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id + 1 }, false);
if (slot.params.cache_prompt) {
slot.cache_tokens.push_back(prompt_tokens[slot.n_past]);
@ -1937,9 +1958,9 @@ struct server_context {
LOG_TEE("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd);
llama_kv_cache_seq_add(ctx, slot.id + 1, slot.ga_i, slot.n_past_se, ib * bd);
llama_kv_cache_seq_div(ctx, slot.id + 1, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n);
llama_kv_cache_seq_add(ctx, slot.id + 1, slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd);
slot.n_past_se -= bd;

379
ggml.c
View File

@ -1841,6 +1841,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"FLASH_ATTN",
"FLASH_FF",
"FLASH_ATTN_BACK",
"SSM_CONV",
"SSM_SCAN",
"WIN_PART",
"WIN_UNPART",
"GET_REL_POS",
@ -1863,7 +1865,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"CROSS_ENTROPY_LOSS_BACK",
};
static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74");
static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@ -1929,6 +1931,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"flash_attn(x)",
"flash_ff(x)",
"flash_attn_back(x)",
"ssm_conv(x)",
"ssm_scan(x)",
"win_part(x)",
"win_unpart(x)",
"get_rel_pos(x)",
@ -1951,7 +1955,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"cross_entropy_loss_back(x,y)",
};
static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74");
static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@ -6154,6 +6158,108 @@ struct ggml_tensor * ggml_flash_attn_back(
return result;
}
// ggml_ssm_conv
struct ggml_tensor * ggml_ssm_conv(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * c,
struct ggml_tensor * sq) {
GGML_ASSERT(ggml_is_3d(s));
GGML_ASSERT(ggml_is_matrix(x));
GGML_ASSERT(ggml_is_matrix(c));
GGML_ASSERT(ggml_is_matrix(sq));
GGML_ASSERT(sq->type == GGML_TYPE_I32);
const int64_t d_conv = c->ne[0];
const int64_t d_inner = c->ne[1];
const int64_t n_tokens = x->ne[1];
const int64_t n_kv = s->ne[2];
GGML_ASSERT( s->ne[0] == d_conv - 1);
GGML_ASSERT( s->ne[1] == d_inner);
GGML_ASSERT( x->ne[0] == d_inner);
GGML_ASSERT(sq->ne[0] == n_kv);
GGML_ASSERT(sq->ne[1] == n_tokens);
bool is_node = false;
if (s->grad || x->grad || c->grad || sq->grad) {
GGML_ASSERT(false); // TODO: implement
is_node = true;
}
// 2-in-1 concatenated x and conv_states, {d_inner, n_tokens} with {d_conv, d_inner, n_kv}
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, (d_inner*n_tokens) + (d_conv*d_inner*n_kv));
result->op = GGML_OP_SSM_CONV;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = s;
result->src[1] = x;
result->src[2] = c;
result->src[3] = sq;
return result;
}
// ggml_ssm_scan
struct ggml_tensor * ggml_ssm_scan(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * dt,
struct ggml_tensor * A,
struct ggml_tensor * B,
struct ggml_tensor * C,
struct ggml_tensor * sq) {
GGML_ASSERT(ggml_is_contiguous(s));
GGML_ASSERT(ggml_is_contiguous(x));
GGML_ASSERT(ggml_is_contiguous(dt));
GGML_ASSERT(ggml_is_contiguous(A));
GGML_ASSERT(sq->type == GGML_TYPE_I32);
GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
GGML_ASSERT(ggml_are_same_shape(x, dt));
{
const int64_t d_state = s->ne[0];
const int64_t d_inner = s->ne[1];
const int64_t n_tokens = x->ne[1];
GGML_ASSERT(x->ne[0] == d_inner);
GGML_ASSERT(A->ne[0] == d_state);
GGML_ASSERT(A->ne[1] == d_inner);
GGML_ASSERT(B->ne[0] == d_state);
GGML_ASSERT(B->ne[1] == n_tokens);
GGML_ASSERT(C->ne[0] == d_state);
GGML_ASSERT(C->ne[1] == n_tokens);
}
bool is_node = false;
if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad || sq->grad) {
GGML_ASSERT(false); // TODO: implement
is_node = true;
}
// 2-in-1 concatenated y and ssm_states, {d_inner, n_tokens} with {d_state, d_inner, n_kv}
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
result->op = GGML_OP_SSM_SCAN;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = s;
result->src[1] = x;
result->src[2] = dt;
result->src[3] = A;
result->src[4] = B;
result->src[5] = C;
result->src[6] = sq;
return result;
}
// ggml_win_part
struct ggml_tensor * ggml_win_part(
@ -14771,6 +14877,257 @@ static void ggml_compute_forward_flash_attn_back(
}
}
// ggml_compute_forward_ssm_conv
static void ggml_compute_forward_ssm_conv_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return;
}
const struct ggml_tensor * src0 = dst->src[0]; // conv_state
const struct ggml_tensor * src1 = dst->src[1]; // x
const struct ggml_tensor * src2 = dst->src[2]; // conv1d.weight
const struct ggml_tensor * src3 = dst->src[3]; // state_seq
const int ith = params->ith;
const int nth = params->nth;
const int nc = src2->ne[0]; // d_conv
const int nr = src0->ne[1]; // d_inner
const int n_t = src1->ne[1]; // n_tokens
const int n_kv = src0->ne[2]; // max number of sequences in the batch
GGML_ASSERT((nr*n_t) + (nc*nr*n_kv) == ggml_nelements(dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
GGML_ASSERT(src2->nb[0] == sizeof(float));
GGML_ASSERT(src3->nb[0] == sizeof(int32_t));
GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
// for use with the destination state offset between sequences
GGML_ASSERT(src2->nb[2] == src2->ne[1]*src2->ne[0]*sizeof(float));
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
const int ir = ir1 - ir0;
if (n_kv > 1) {
// multiple sequences means it's hard to know when it's the first time a state is read,
// so copy them all over to the destination, just to be sure.
for (int i3 = 0; i3 < n_kv; ++i3) {
float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]));
float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + i3*(src2->nb[2]) + nr*n_t*sizeof(float));
// can't use memcpy because of d_conv vs d_conv - 1
for (int i1 = 0; i1 < ir; ++i1) {
for (int i0 = 0; i0 < nc - 1; ++i0) {
// copy s0 to last (d_conv - 1) columns of s
s[1 + i0 + i1*nc] = s0[i0 + i1*(nc - 1)];
}
}
}
}
for (int i2 = 0; i2 < n_t; ++i2) {
int32_t * sq = (int32_t *) ((char *) src3->data + i2*(src3->nb[1])); // {n_kv, n_tokens}
float * x = (float *) ((char *) dst->data + ir0*sizeof(float) + i2*(nr*sizeof(float))); // {d_inner, n_tokens}
float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + sq[0]*(src2->nb[2]) + nr*n_t*sizeof(float)); // {d_conv, d_inner, n_kv}
float * s0; // {d_conv - 1, d_inner, n_kv}
float * x0 = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens}
float * c = (float *) ((char *) src2->data + ir0*(src2->nb[1])); // {d_conv, d_inner}
int ne0s0;
GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv);
// avoid needing to copy the state for the first token
if (i2 == 0) {
s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_conv - 1, d_inner, n_kv}
ne0s0 = src0->ne[0];
} else {
// the source is the last (d_conv - 1) columns of the destination
s0 = s + 1;
ne0s0 = nc;
}
// d_inner
for (int i1 = 0; i1 < ir; ++i1) {
// shift state left
for (int i0 = 0; i0 < nc - 1; ++i0) {
s[i0 + i1*nc] = s0[i0 + i1*ne0s0];
}
// insert x on the last column
s[(nc - 1) + i1*nc] = x0[i1];
}
// handle copies when there are multiple output states
for (int i3 = 1; i3 < n_kv; ++i3) {
int32_t seq = sq[i3];
if (0 <= seq && seq < n_kv) {
float * s1 = s + (seq - sq[0])*nc*nr;
memcpy(s1, s, nc*ir*sizeof(float));
} else {
// stop at negative or too big seq_ids
break;
}
}
// it seems a little faster when this is separate from the state shift
for (int i1 = 0; i1 < ir; ++i1) {
// rowwise dot product
float sumf = 0.0f;
for (int i0 = 0; i0 < nc; ++i0) {
int i = i0 + i1*nc;
sumf += s[i] * c[i];
}
x[i1] = sumf;
}
}
}
static void ggml_compute_forward_ssm_conv(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
switch (dst->src[0]->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_ssm_conv_f32(params, dst);
} break;
default:
{
GGML_ASSERT(false);
} break;
}
}
// ggml_compute_forward_ssm_scan
static void ggml_compute_forward_ssm_scan_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return;
}
const struct ggml_tensor * src0 = dst->src[0]; // s
const struct ggml_tensor * src1 = dst->src[1]; // x
const struct ggml_tensor * src2 = dst->src[2]; // dt
const struct ggml_tensor * src3 = dst->src[3]; // A
const struct ggml_tensor * src4 = dst->src[4]; // B
const struct ggml_tensor * src5 = dst->src[5]; // C
const struct ggml_tensor * src6 = dst->src[6]; // sq
const int ith = params->ith;
const int nth = params->nth;
const int64_t nc = src0->ne[0]; // d_state
const int64_t nr = src0->ne[1]; // d_inner
const int64_t n_t = src1->ne[1]; // number of tokens in the batch
const int64_t n_kv = src0->ne[2]; // max number of sequences in the batch
GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT(src1->nb[0] == sizeof(float));
GGML_ASSERT(src2->nb[0] == sizeof(float));
GGML_ASSERT(src3->nb[0] == sizeof(float));
GGML_ASSERT(src4->nb[0] == sizeof(float));
GGML_ASSERT(src5->nb[0] == sizeof(float));
// required for the dot product between s and C, and when copying the states
GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
// required for per-sequence offsets for states
GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float));
// required to get correct offset for state destination (i.e. src1->nb[2])
GGML_ASSERT(src1->nb[2] == src1->ne[0]*src1->ne[1]*sizeof(float));
// rows per thread
const int dr = (nr + nth - 1)/nth;
// row range for this thread
const int ir0 = dr*ith;
const int ir1 = MIN(ir0 + dr, nr);
const int ir = ir1 - ir0;
if (n_kv > 1) {
// it's hard to know if the source states have already been copied
// when there are multiple, so copy them already.
for (int i3 = 0; i3 < n_kv; ++i3) {
float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]));
float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[2]);
memcpy(s, s0, nc*ir*sizeof(float));
}
}
for (int i2 = 0; i2 < n_t; ++i2) {
int32_t * sq = (int32_t *) ((char *) src6->data + i2*(src6->nb[1])); // {n_kv, n_tokens}
float * y = (float *) ((char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens}
float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2]) + src1->nb[2]); // {d_state, d_inner, n_kv}
float * s0;
float * x = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens}
float * dt = (float *) ((char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1])); // {d_inner, n_tokens}
float * A = (float *) ((char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
float * B = (float *) ((char *) src4->data + i2*(src4->nb[1])); // {d_state, n_tokens}
float * C = (float *) ((char *) src5->data + i2*(src5->nb[1])); // {d_state, n_tokens}
GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv);
// avoid needing to copy the state for the first token
if (i2 == 0) {
s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_state, d_inner, n_kv}
} else {
// otherwise the source is the same as the destination
s0 = s;
}
// d_inner
for (int i1 = 0; i1 < ir; ++i1) {
// ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
float x_dt = x[i1] * dt_soft_plus;
float sumf = 0.0f;
// d_state
for (int i0 = 0; i0 < nc; ++i0) {
int i = i0 + i1*nc;
// state = prev_state * dA + dB * x
float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
// y = rowwise_dotprod(state, C)
sumf += state * C[i0];
s[i] = state;
}
y[i1] = sumf;
}
// handle copies when there are multiple output states
for (int i3 = 1; i3 < n_kv; ++i3) {
int32_t seq = sq[i3];
if (0 <= seq && seq < n_kv) {
float * s1 = s + (seq - sq[0])*nc*nr;
memcpy(s1, s, nc*ir*sizeof(float));
} else {
// stop at negative or too big seq_ids
break;
}
}
}
}
static void ggml_compute_forward_ssm_scan(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
switch (dst->src[0]->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_ssm_scan_f32(params, dst);
} break;
default:
{
GGML_ASSERT(false);
} break;
}
}
// ggml_compute_forward_win_part
static void ggml_compute_forward_win_part_f32(
@ -15830,6 +16187,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
bool masked = t != 0;
ggml_compute_forward_flash_attn_back(params, masked, tensor);
} break;
case GGML_OP_SSM_CONV:
{
ggml_compute_forward_ssm_conv(params, tensor);
} break;
case GGML_OP_SSM_SCAN:
{
ggml_compute_forward_ssm_scan(params, tensor);
} break;
case GGML_OP_WIN_PART:
{
ggml_compute_forward_win_part(params, tensor);
@ -16884,6 +17249,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
{
GGML_ASSERT(false); // not supported
} break;
case GGML_OP_SSM_CONV:
case GGML_OP_SSM_SCAN:
{
GGML_ASSERT(false); // TODO: not implemented
} break;
case GGML_OP_WIN_PART:
case GGML_OP_WIN_UNPART:
case GGML_OP_UNARY:
@ -17590,6 +17960,11 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
{
n_tasks = n_threads;
} break;
case GGML_OP_SSM_CONV:
case GGML_OP_SSM_SCAN:
{
n_tasks = n_threads;
} break;
case GGML_OP_WIN_PART:
case GGML_OP_WIN_UNPART:
case GGML_OP_GET_REL_POS:

19
ggml.h
View File

@ -472,6 +472,8 @@ extern "C" {
GGML_OP_FLASH_ATTN,
GGML_OP_FLASH_FF,
GGML_OP_FLASH_ATTN_BACK,
GGML_OP_SSM_CONV,
GGML_OP_SSM_SCAN,
GGML_OP_WIN_PART,
GGML_OP_WIN_UNPART,
GGML_OP_GET_REL_POS,
@ -1728,6 +1730,23 @@ extern "C" {
struct ggml_tensor * c0,
struct ggml_tensor * c1);
GGML_API struct ggml_tensor * ggml_ssm_conv(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * c,
struct ggml_tensor * sq);
GGML_API struct ggml_tensor * ggml_ssm_scan(
struct ggml_context * ctx,
struct ggml_tensor * s,
struct ggml_tensor * x,
struct ggml_tensor * dt,
struct ggml_tensor * A,
struct ggml_tensor * B,
struct ggml_tensor * C,
struct ggml_tensor * sq);
// partition into non-overlapping windows with padding if needed
// example:
// a: 768 64 64 1

41
gguf-py/gguf/constants.py
View File

@ -61,6 +61,12 @@ class Keys:
SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length"
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
class SSM:
CONV_KERNEL = "{arch}.ssm.conv_kernel"
INNER_SIZE = "{arch}.ssm.inner_size"
STATE_SIZE = "{arch}.ssm.state_size"
TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
class Tokenizer:
MODEL = "tokenizer.ggml.model"
LIST = "tokenizer.ggml.tokens"
@ -113,6 +119,7 @@ class MODEL_ARCH(IntEnum):
MINICPM = auto()
GEMMA = auto()
STARCODER2 = auto()
MAMBA = auto()
class MODEL_TENSOR(IntEnum):
@ -144,6 +151,13 @@ class MODEL_TENSOR(IntEnum):
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
SSM_IN = auto()
SSM_CONV1D = auto()
SSM_X = auto()
SSM_DT = auto()
SSM_A = auto()
SSM_D = auto()
SSM_OUT = auto()
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@ -171,6 +185,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.MAMBA: "mamba",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@ -202,6 +217,13 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",
MODEL_TENSOR.SSM_DT: "blk.{bid}.ssm_dt",
MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a",
MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@ -543,6 +565,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.MAMBA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.SSM_IN,
MODEL_TENSOR.SSM_CONV1D,
MODEL_TENSOR.SSM_X,
MODEL_TENSOR.SSM_DT,
MODEL_TENSOR.SSM_A,
MODEL_TENSOR.SSM_D,
MODEL_TENSOR.SSM_OUT,
],
# TODO
}
@ -734,6 +769,12 @@ KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR
KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN
KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED
# SSM
KEY_SSM_CONV_KERNEL = Keys.SSM.CONV_KERNEL
KEY_SSM_INNER_SIZE = Keys.SSM.INNER_SIZE
KEY_SSM_STATE_SIZE = Keys.SSM.STATE_SIZE
KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK
# tokenization
KEY_TOKENIZER_MODEL = Keys.Tokenizer.MODEL
KEY_TOKENIZER_LIST = Keys.Tokenizer.LIST

12
gguf-py/gguf/gguf_writer.py
View File

@ -382,6 +382,18 @@ class GGUFWriter:
def add_rope_scaling_finetuned(self, value: bool) -> None:
self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value)
def add_ssm_conv_kernel(self, value: int) -> None:
self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value)
def add_ssm_inner_size(self, value: int) -> None:
self.add_uint32(Keys.SSM.INNER_SIZE.format(arch=self.arch), value)
def add_ssm_state_size(self, value: int) -> None:
self.add_uint32(Keys.SSM.STATE_SIZE.format(arch=self.arch), value)
def add_ssm_time_step_rank(self, value: int) -> None:
self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value)
def add_tokenizer_model(self, model: str) -> None:
self.add_string(Keys.Tokenizer.MODEL, model)

46
gguf-py/gguf/tensor_mapping.py
View File

@ -20,6 +20,9 @@ class TensorNameMap:
"wte", # gpt2
"transformer.embd.wte", # phi2
"model.tok_embeddings", # internlm2
"model.embedding", # mamba-qbert
"backbone.embedding", # mamba
"backbone.embeddings", # mamba-hf
),
# Token type embeddings
@ -44,7 +47,7 @@ class TensorNameMap:
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
@ -61,6 +64,8 @@ class TensorNameMap:
"language_model.encoder.final_layernorm", # persimmon
"model.final_layernorm", # persimmon
"lm_head.ln", # phi2
"model.norm_f", # mamba-qbert
"backbone.norm_f", # mamba
),
# Rope frequencies
@ -86,6 +91,8 @@ class TensorNameMap:
"transformer.h.{bid}.ln", # phi2
"model.layers.layers.{bid}.norm", # plamo
"model.layers.{bid}.attention_norm", # internlm2
"model.layers.{bid}.norm", # mamba-qbert
"backbone.layers.{bid}.norm", # mamba
),
# Attention norm 2
@ -282,7 +289,42 @@ class TensorNameMap:
MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
)
),
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",
),
}
mapping: dict[str, tuple[MODEL_TENSOR, str]]

698
llama.cpp
View File

File diff suppressed because it is too large Load Diff

4
llama.h
View File

@ -235,6 +235,7 @@ extern "C" {
uint32_t seed; // RNG seed, -1 for random
uint32_t n_ctx; // text context, 0 = from model
uint32_t n_batch; // prompt processing maximum batch size
uint32_t n_parallel; // number of parallel sequences (i.e. distinct states for recurrent models)
uint32_t n_threads; // number of threads to use for generation
uint32_t n_threads_batch; // number of threads to use for batch processing
@ -376,6 +377,7 @@ extern "C" {
LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_max_seq (const struct llama_context * ctx);
LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model);
@ -502,7 +504,7 @@ extern "C" {
// seq_id < 0 : match any sequence
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_rm(
LLAMA_API bool llama_kv_cache_seq_rm(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,