llama.cpp/examples/batched-bench/batched-bench.cpp

249 lines
7.4 KiB
C++
Raw Normal View History

#include "common.h"
#include "llama.h"
#include <algorithm>
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
// mutates the input string
static std::vector<int> parse_list(char * p) {
std::vector<int> ret;
char * q = p;
while (*p) {
if (*p == ',') {
*p = '\0';
ret.push_back(std::atoi(q));
q = p + 1;
}
++p;
}
ret.push_back(std::atoi(q));
return ret;
}
int main(int argc, char ** argv) {
gpt_params params;
if (argc == 1 || argv[1][0] == '-') {
printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
printf(" <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
printf(" example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
return 1 ;
}
int n_kv_max = 2048;
int is_pp_shared = 0;
int n_gpu_layers = 0;
std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
std::vector<int> n_tg = { 128, 256, };
std::vector<int> n_pl = { 1, 2, 4, 8, 16, 32, };
//std::vector<int> n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, };
if (argc >= 2) {
params.model = argv[1];
}
if (argc >= 3) {
n_kv_max = std::atoi(argv[2]);
}
if (argc >= 4) {
is_pp_shared = std::atoi(argv[3]);
}
if (argc >= 5) {
n_gpu_layers = std::atoi(argv[4]);
}
if (argc >= 6) {
n_pp = parse_list(argv[5]);
}
if (argc >= 7) {
n_tg = parse_list(argv[6]);
}
if (argc >= 8) {
n_pl = parse_list(argv[7]);
}
// init LLM
ggml : add numa options (#5377) * Added numa options to allow finer grained control as well as plumbing for a new mirror mode that will require numa.h * Reverted Makefile * Fixed include * Removed sched.h from ggml.h, moved ggml_get_numa_affinity into ggml.c, removed trailing whitespace and fixed up a few inconsistent variables * removed trailing whitespace * Added numa options to allow finer grained control as well as plumbing for a new mirror mode that will require numa.h * Reverting Makefile * Fixed a number of issues with the move from BOOL to ggml_numa_strategies. Added a note about mirror mode note being implemented yet * Removing MIRROR_MODE code for this PR * Removing last bit of MIRROR_MODE code for this PR * Removing unneeded branch in server.cpp example and moving get_numa_affinity and making it static * Fixed lingering init_llama_backend() bool calls in tests and examples * Remote enum llama_numa_strategies * Revert bad merge with dynatemp flags * add missing enum ggml_numa_strategies declaration and revert sync problem with master * add missing enum ggml_numa_strategies declaration * fixed ggml_init_numa variable * Update ggml.h Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * Update READMEs with info about numa flags, change INTERLEAVE strategy name to DISTRIBUTE everywhere, implement the improved distribution strategy from @rankaiyx, fix a spelling mistake and un-merge some bad merges * split numa init out from llama_backend_init and created llama_numa_init. Updated all code paths and samples * Fix up some boolean vs enum comparisons * Added #ifdefs for non-Linux OS that don't have cpu_set_t datatype * Update ggml.h Align enum values Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update ggml.c Remove whitespace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update ggml.c align paremeters Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update examples/server/server.cpp remove whitespace and align brace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Update common/common.cpp Remove whitespace and align brace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * unified ggml_numa_strategy enum and fixed text alignment in server.cpp example * Update ggml.c simplified return for platforms without NUMA support Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> * removed redundant else from cli argument processing of --numa * whitespace --------- Co-authored-by: root <root@nenya.lothlorien.ca> Co-authored-by: Jared Van Bortel <cebtenzzre@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Jared Van Bortel <jared@nomic.ai>
2024-02-16 09:31:07 +00:00
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model
llama_model_params model_params = llama_model_default_params();
const std::vector<float> t_split(llama_max_devices(), 0.0f);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
model_params.n_gpu_layers = n_gpu_layers;
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 19:07:38 +00:00
model_params.tensor_split = t_split.data();
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
llama_context_params ctx_params = llama_context_default_params();
ctx_params.seed = 1234;
ctx_params.n_ctx = n_kv_max;
ctx_params.n_batch = 512;
ctx_params.n_threads = params.n_threads;
ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
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 22:31:00 +00:00
// 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) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
// decode in batches of ctx_params.n_batch tokens
auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
}
return true;
};
// warm up
{
for (int i = 0; i < 16; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
}
2023-10-25 07:26:27 +00:00
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
2023-10-25 07:26:27 +00:00
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
for ( int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
for ( int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
for (int i_pl = 0; i_pl < (int) n_pl.size(); ++i_pl) {
const int pp = n_pp[i_pp];
const int tg = n_tg[i_tg];
const int pl = n_pl[i_pl];
const int n_ctx_req = is_pp_shared ? pp + pl*tg : pl*(pp + tg);
if (n_ctx_req > n_kv_max) {
continue;
}
llama_batch_clear(batch);
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 22:31:00 +00:00
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;
const auto t_pp_start = ggml_time_us();
llama_kv_cache_clear(ctx);
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
if (is_pp_shared) {
for (int32_t i = 1; i < pl; ++i) {
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 22:31:00 +00:00
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
}
const auto t_pp_end = ggml_time_us();
const auto t_tg_start = ggml_time_us();
for (int i = 0; i < tg; ++i) {
llama_batch_clear(batch);
for (int j = 0; j < pl; ++j) {
llama_batch_add(batch, 0, pp + i, { j }, true);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
return 1;
}
}
const auto t_tg_end = ggml_time_us();
const int32_t n_kv = n_ctx_req;
const float t_pp = (t_pp_end - t_pp_start) / 1000000.0f;
const float t_tg = (t_tg_end - t_tg_start) / 1000000.0f;
const float t = t_pp + t_tg;
const float speed_pp = is_pp_shared ? pp / t_pp : pl*pp / t_pp;
const float speed_tg = pl*tg / t_tg;
const float speed = n_kv / t;
LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
}
}
}
llama_print_timings(ctx);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
return 0;
}