mirror of
https://git.adityakumar.xyz/llama.cpp.git
synced 2024-11-09 15:29:43 +00:00
ggml : implement backward pass for llama + small training-llama-from-scratch example (#1360)
* implement 8 of 14 missing backward pass operations used by llama - GGML_OP_ADD_AT - GGML_OP_CPY - GGML_OP_MUL_MAT (src0.grad) - GGML_OP_PERMUTE - GGML_OP_RESHAPE - GGML_OP_SCALE - GGML_OP_TRANSPOSE - GGML_OP_VIEW implement additional ggml operation GGML_OP_ADD_AT, which is necessary for backward pass of GGML_OP_VIEW. this operation adds src1 to src0 with data offset, i.e. to view(src0, ..., offset). the values are return in a tensor size of src0. values outside of [data+offset:data+offset+nbytes(src1)] are just the original values from src0. still missing backward passes for llama: - GGML_OP_DIAG_MASK_INF - GGML_OP_GET_ROWS - GGML_OP_RMS_NORM - GGML_OP_ROPE - GGML_OP_SILU - GGML_OP_SOFT_MAX * implement 5 of 6 missing backward pass operations used by llama - GGML_OP_DIAG_MASK_INF - GGML_OP_GET_ROWS - GGML_OP_RMS_NORM - GGML_OP_SILU - GGML_OP_SOFT_MAX add necessary ggml operations GGML_OP_ADD1, GGML_OP_SILU_BACK, GGML_OP_RMS_NORM_BACK, GGML_OP_DIAG_MASK_ZERO, and GGML_OP_ROPE_BACK GGML_OP_ADD1 is necessary to add a scalar value in the backward pass of GGML_OP_SOFT_MAX GGML_OP_ADD1 could also be replaced by using GGML_OP_ADD and GGML_OP_REPEAT, but the performance would be worse. additionally GGML_OP_REPEAT will return unexpected value when the the input to GGML_OP_SOFT_MAX contains only a single scalar. in this case GGML_OP_REPEAT will not return the value that should be repeated (src1) but the value which shape the result should take (src0). So in this case it can not replace GGML_OP_ADD1. GGML_OP_SILU_BACK, GGML_OP_RMS_NORM_BACK and GGML_OP_ROPE_BACK are necessary for backward pass of GGML_OP_SILU, GGML_OP_RMS_NORM and GGML_OP_ROPE. The backward pass for these functions cannot be easily composed of existing operations. Since the backward pass builds a computation graph we need operations forward pass implementations of the the required backward passes. Sounds a bit confusing at first, I know... GGML_OP_DIAG_MASK_ZERO is necessary for backward pass of GGML_OP_DIAG_MASK_INF. Some operations where previously inplace-only. for backward pass there needs to be non-inplace variants. staying consistent with other operations that have non-inplace and inplace variants, the operations are changed to non-inplace and functions with "_inplace" are added which are inplace. in llama we need to call the inplace variants so that it is implemented as before. for llama backward pass we need to use the non-inplace variants. still not completely implemented backward passes for llama: - GGML_OP_ROPE: needs forward pass for GGML_OP_ROPE_BACK - GGML_OP_GET_ROWS: only necessary for tokenizer * norm & rms_norm can not be threaded: after investigation rms norm for quite some time I come to the conclusion that neither norm, nor rms_norm can be threaded, because we need mean over all items, not just of the slices each thread sees. * remove already resolved TODO * implement backward pass of ggml_rope and ggml_rope_back * implement backward pass for ggml_get_rows and for new operation ggml_get_rows_back * add test-grad0.c * use GGML_PRINT_DEBUG for debug messages which will otherwise flood the console * test both gradients of mul_mat * disable graph dot export as it floods console * bug fixes for silu_back * successfully test silu backward * bug fix for scale backward pass use sum instead of mean for gradient of scalar scale parameter * successfully test scale backward * improve performance of sum backward pass use add1(x,y) instead of add(x,repeat(y,x)) * improve performance of sqr backward pass use scale(x,y) instead of mul(x,repeat(y,x)) * successfully test rope backward * bug fix for cpy backward pass * successfully test cpy backward * bug fix for reshape backward pass * successfully test reshape backward * add test-opt.c this uses ggml_opt to train a,b for minimal e=sum(sqr(c - a*b)) for random initial a,b,c * correctly implement softmax backward pass using new operation ggml_diag ggml_diag constructs diagonal matrices with entries. ggml_diag(shape[a,1,c,d]) -> shape[a,a,c,d] * successfully test soft_max backward * align shape annotations * add shape annotations for llama * de-duplicate ggml_forward_dup code taking care of contiguous tensors of same type. with this we can duplicate tensor of any typ as long as they are contiguous. * fix ggml_compute_forward_dup_same_cont for when nelements < nthreads when more threads are used than elements exist ie1 was less than ie0, resulting in invalid negative byte count argument in memcpy * bug fix for add_at forward required for view backward pass src0 values must be copied to dst, because during addition we don't touch all dst elements in contrast to the normal add function. * successfully test view backward * minor code format improvement * fix ggml_forward_add functions to work correctly with transposed tensors uses the same logic as in ggml_compute_forward_add_q_f32, but make it consistent across all ggml_compute_forward_add_... functions. this also slightly changes the mem access pattern of the different threads to works as in ggml_compute_forward_add_q_f32. * fix ggml_forward_add1 functions to work correctly with transposed tensors uses the same logic as in ggml_compute_forward_add1_q_f32, but make it consistent across all ggml_compute_forward_add1_... functions. this also slightly changes the mem access pattern of the different threads to works as in ggml_compute_forward_add1_q_f32. * test-grad0.c : add print_elements to help with debugging * successfully test permute backward * some minor test-grad0 fixes * fix sub, mul and div functions to work correctly with transposed tensors uses the same logic as in add * implement ggml_cont backward pass * successfully test transpose backward and permute for all permutations also test sub, mul and div up to max n_dims * test-grad0.c add TODO for view_2d and view_3d add_at (required for view backward pass) is a bit tricky for n_dims > 1. * fix comments * successfully test diag_mask_inf and diag_mask_zero backward * test-grad0 : fix test for div nargs and ndims was swapped, corrupting the stack * fix diag_mask to work with non-inplace input * move dup call into the actual add_at functions * fix get rows backward pass * successfully test get_rows backward * fix view backward pass add nb parameters to add_at like in view. together with offset they define how to view dst and src0 during the add_at operation. * successfully test backward pass of view_1d, view_2d and view_3d * fix backward pass for rms_norm I would have used formulas from other frameworks, but they differed so I could not decide which is correct. Instead it was derived here in comment using manual forward-backward automatic differention of rms_norm and simplification. * successfully test backward pass of rms_norm some tests may fail when gradients are large. could not find a satisfying configuration to check for abs error and relative error that passes all tests while still actually testing the results with tight enough error bounds. when looking at the values the "failed" tests look actually ok. for example: rms_norm: ndims=2, i=0, k=2, x0=0.000153, xm=0.000053, xp=0.000253, f0=0.278594, f1=0.086213, g0=961.905457, g1=966.064941, eps=0.000100, error_abs=4.159485, error_rel=0.004324 it is due to the test logic in check_gradients that they fail. * add todos for llama backward pass - implementation for ADD1 backward pass should probably use sum instead of mean (but this backward pass is not required) - repeat is not yet tested and looks like it only works for single element src0 inputs. * add operation ggml_sum_rows ggml_sum_rows(shape[a,b,c,d]) -> shape[1,b,c,d] * add missing GGML_OP_SUM_ROWS * fix backward pass for repeat requires ggml_sum_rows * successfully test backward pass of repeat * update quantization types in switch-case of add_at and add1 * add baby-llama example training a very small llama model from scratch to output a sinusoidal wave. had to increase maximum number of optimization parameters to train from scratch. * fix softmax in baby-llama example * switching from training with adam to lbfgs produces much better results in the baby-llama example * train with two examples, creating new tensors each time.. * fix bug when using ggml_opt to optimize params in one context and use a renewable context for eval and opt when not keeping gradients of model parameters they are overwritten by tensors created by opt, which may be invalid after opt context is renewed. so we need to keep the original gradients and make dups for opt * train on multiple examples, generate & print tokens with trained model afterwards ctx0 for evaluation and optimization is renewed for each sample * add ggml_reshape_1d, ggml_reshape_4d and ggml_view_4d * fix soft_max backward pass for input->ne[1] != 1 * add ggml_log operation necessary for cross entropy loss * add test for ggml_log gradients * implement backward pass for ggml_sum_rows, necessary for cross entropy loss * implement ggml_repeat support for rank > 2 tensors * add test for ggml_sum_rows gradients * fix training get_example_targets predict the next token, not the current token! * add square_error_loss and cross_entropy_loss functions * optimize loss over multiple samples this increases computation graph, need parallel batched forward for more efficiency. * fix backward pass for add_at and change arguments to have same order as in view * add ggml_set(ctx, a, b) to set b in view of a and return modified a necessary to set values into kv_self cache and properly propagate the gradients * fix kv_self gradients for training use ggml_set instead of ggml_cpy to set kv_self cache with properly propagating gradients * replace inplace operations for training with copying operations to allow gradient propagation * add GGML_ASSERT to catch ggml_rope and back value errors * add trainable lora-only model with all big matrices C split into A,B with A*B=C this is not a lora-finetune, but the whole model changed to have only low-rank "lora" matrices. training this instead of the normal model resulted in much worse results though... * vastly improve training results instead of logit targets 0 and 1 use -1 and +1. * shorten code using a variable * change name of GGML_OP_ADD_AT to GGML_OP_ACC * smaller default values for baby llama model parameters * update static assert of GGML_OP_COUNT * remove shape annotations in llama_eval_internal * revert disabling of threading for rms_norm and norm * rename print functions in baby-llama example * fix call to ggml_set_name * add missing include for strcmp, etc * remove trailing whitespace * reduce number of test-grad0 iterations avoid exceeding timeout of automated tests * remove busy loop that was used as sleep for slower sinus wave generation * disable slow tests grad0 and opt to avoid exceeding timeouts * c++ in baby-llama example use c++ includes instead of c includes use std::min, std::max instead of MIN, MAX macros * c++ in baby-llama example use c++ includes instead of c includes use std::min, std::max instead of MIN, MAX macros * ggml : fix compiler warnings + cosmetic changes * ggml : fix nullptr derefs in GGML_OP_CONT and GGML_OP_RESHAPE back * swap arguments to vDSP_vdiv call documentation for vDSP_vdiv states: "Note that B comes before A!" * swap arguments to vDSP_vdiv call documentation for vDSP_vdiv states: "Note that B comes before A!" * ggml : swap vDSP_vsub args as per documentation * add parallel batched forward function for baby-llama training * cleanup code for batched training * remove trailing whitespace * minor : fix compiler warnings + indentation style * ggml : fix null ptr deref in backward pass * ggml : remove Q4_2 remnants * ggml : fix clang-tidy warnings * baby-llama : couple of clang-tidy warnings --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
xaedes
GitHub
parent
f048af0230
commit
f954edda93
9 changed files with 6157 additions and 261 deletions
|
|
@ -36,4 +36,5 @@ else()
|
|||
add_subdirectory(embedding)
|
||||
add_subdirectory(save-load-state)
|
||||
add_subdirectory(benchmark)
|
||||
add_subdirectory(baby-llama)
|
||||
endif()
|
||||
|
|
|
|||
4
examples/baby-llama/CMakeLists.txt
Normal file
4
examples/baby-llama/CMakeLists.txt
Normal file
|
|
@ -0,0 +1,4 @@
|
|||
set(TARGET baby-llama)
|
||||
add_executable(${TARGET} baby-llama.cpp)
|
||||
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
|
||||
target_compile_features(${TARGET} PRIVATE cxx_std_11)
|
||||
1687
examples/baby-llama/baby-llama.cpp
Normal file
1687
examples/baby-llama/baby-llama.cpp
Normal file
File diff suppressed because it is too large
Load diff
200
ggml.h
200
ggml.h
|
|
@ -192,7 +192,7 @@
|
|||
|
||||
#define GGML_MAX_DIMS 4
|
||||
#define GGML_MAX_NODES 4096
|
||||
#define GGML_MAX_PARAMS 16
|
||||
#define GGML_MAX_PARAMS 256
|
||||
#define GGML_MAX_CONTEXTS 64
|
||||
#define GGML_MAX_OPT 4
|
||||
#define GGML_DEFAULT_N_THREADS 4
|
||||
|
|
@ -262,12 +262,16 @@ extern "C" {
|
|||
|
||||
GGML_OP_DUP,
|
||||
GGML_OP_ADD,
|
||||
GGML_OP_ADD1,
|
||||
GGML_OP_ACC,
|
||||
GGML_OP_SUB,
|
||||
GGML_OP_MUL,
|
||||
GGML_OP_DIV,
|
||||
GGML_OP_SQR,
|
||||
GGML_OP_SQRT,
|
||||
GGML_OP_LOG,
|
||||
GGML_OP_SUM,
|
||||
GGML_OP_SUM_ROWS,
|
||||
GGML_OP_MEAN,
|
||||
GGML_OP_REPEAT,
|
||||
GGML_OP_ABS,
|
||||
|
|
@ -277,12 +281,15 @@ extern "C" {
|
|||
GGML_OP_RELU,
|
||||
GGML_OP_GELU,
|
||||
GGML_OP_SILU,
|
||||
GGML_OP_SILU_BACK,
|
||||
GGML_OP_NORM, // normalize
|
||||
GGML_OP_RMS_NORM,
|
||||
GGML_OP_RMS_NORM_BACK,
|
||||
|
||||
GGML_OP_MUL_MAT,
|
||||
|
||||
GGML_OP_SCALE,
|
||||
GGML_OP_SET,
|
||||
GGML_OP_CPY,
|
||||
GGML_OP_CONT,
|
||||
GGML_OP_RESHAPE,
|
||||
|
|
@ -290,9 +297,13 @@ extern "C" {
|
|||
GGML_OP_PERMUTE,
|
||||
GGML_OP_TRANSPOSE,
|
||||
GGML_OP_GET_ROWS,
|
||||
GGML_OP_GET_ROWS_BACK,
|
||||
GGML_OP_DIAG,
|
||||
GGML_OP_DIAG_MASK_INF,
|
||||
GGML_OP_DIAG_MASK_ZERO,
|
||||
GGML_OP_SOFT_MAX,
|
||||
GGML_OP_ROPE,
|
||||
GGML_OP_ROPE_BACK,
|
||||
GGML_OP_ALIBI,
|
||||
GGML_OP_CONV_1D_1S,
|
||||
GGML_OP_CONV_1D_2S,
|
||||
|
|
@ -496,6 +507,29 @@ extern "C" {
|
|||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_add1(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_acc(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t nb2,
|
||||
size_t nb3,
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_acc_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t nb2,
|
||||
size_t nb3,
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_sub(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
|
|
@ -519,12 +553,24 @@ extern "C" {
|
|||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_log(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_log_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// return scalar
|
||||
// TODO: compute sum along rows
|
||||
GGML_API struct ggml_tensor * ggml_sum(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
|
||||
GGML_API struct ggml_tensor * ggml_sum_rows(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// mean along rows
|
||||
GGML_API struct ggml_tensor * ggml_mean(
|
||||
struct ggml_context * ctx,
|
||||
|
|
@ -566,6 +612,13 @@ extern "C" {
|
|||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// a - x
|
||||
// b - dy
|
||||
GGML_API struct ggml_tensor * ggml_silu_back(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
// normalize along rows
|
||||
// TODO: eps is hardcoded to 1e-5 for now
|
||||
GGML_API struct ggml_tensor * ggml_norm(
|
||||
|
|
@ -576,6 +629,13 @@ extern "C" {
|
|||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// a - x
|
||||
// b - dy
|
||||
GGML_API struct ggml_tensor * ggml_rms_norm_back(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
// A: m rows, n columns
|
||||
// B: p rows, n columns (i.e. we transpose it internally)
|
||||
// result is m columns, p rows
|
||||
|
|
@ -588,12 +648,66 @@ extern "C" {
|
|||
// operations on tensors without backpropagation
|
||||
//
|
||||
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_scale(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_scale_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
// b -> view(a,offset,nb1,nb2,3), return modified a
|
||||
GGML_API struct ggml_tensor * ggml_set(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t nb2,
|
||||
size_t nb3,
|
||||
size_t offset);
|
||||
|
||||
// b -> view(a,offset,nb1,nb2,3), return view(a)
|
||||
GGML_API struct ggml_tensor * ggml_set_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t nb2,
|
||||
size_t nb3,
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_set_1d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_set_1d_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t offset);
|
||||
|
||||
// b -> view(a,offset,nb1,nb2,3), return modified a
|
||||
GGML_API struct ggml_tensor * ggml_set_2d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t offset);
|
||||
|
||||
// b -> view(a,offset,nb1,nb2,3), return view(a)
|
||||
GGML_API struct ggml_tensor * ggml_set_2d_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
size_t nb1,
|
||||
size_t offset);
|
||||
|
||||
|
||||
// a -> b, return view(b)
|
||||
GGML_API struct ggml_tensor * ggml_cpy(
|
||||
struct ggml_context * ctx,
|
||||
|
|
@ -614,6 +728,11 @@ extern "C" {
|
|||
|
||||
// return view(a)
|
||||
// TODO: when we start computing gradient, make a copy instead of view
|
||||
GGML_API struct ggml_tensor * ggml_reshape_1d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int64_t ne0);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_reshape_2d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
|
|
@ -629,6 +748,14 @@ extern "C" {
|
|||
int64_t ne1,
|
||||
int64_t ne2);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_reshape_4d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int64_t ne0,
|
||||
int64_t ne1,
|
||||
int64_t ne2,
|
||||
int64_t ne3);
|
||||
|
||||
// offset in bytes
|
||||
GGML_API struct ggml_tensor * ggml_view_1d(
|
||||
struct ggml_context * ctx,
|
||||
|
|
@ -654,6 +781,18 @@ extern "C" {
|
|||
size_t nb2, // slice stride in bytes
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_view_4d(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int64_t ne0,
|
||||
int64_t ne1,
|
||||
int64_t ne2,
|
||||
int64_t ne3,
|
||||
size_t nb1, // row stride in bytes
|
||||
size_t nb2, // slice stride in bytes
|
||||
size_t nb3,
|
||||
size_t offset);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_permute(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
|
|
@ -672,20 +811,50 @@ extern "C" {
|
|||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_get_rows_back(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
struct ggml_tensor * c);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_diag(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// set elements above the diagonal to -INF
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_diag_mask_inf(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past);
|
||||
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past);
|
||||
|
||||
// set elements above the diagonal to 0
|
||||
GGML_API struct ggml_tensor * ggml_diag_mask_zero(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past);
|
||||
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * gml_diag_mask_zero_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_soft_max(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// rotary position embedding
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_soft_max_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a);
|
||||
|
||||
// rotary position embedding
|
||||
// if mode & 1 == 1, skip n_past elements
|
||||
// if mode & 2 == 1, GPT-NeoX style
|
||||
// TODO: avoid creating a new tensor every time
|
||||
|
|
@ -696,6 +865,23 @@ extern "C" {
|
|||
int n_dims,
|
||||
int mode);
|
||||
|
||||
// in-place, returns view(a)
|
||||
GGML_API struct ggml_tensor * ggml_rope_inplace(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past,
|
||||
int n_dims,
|
||||
int mode);
|
||||
|
||||
// rotary position embedding backward, i.e compute dx from dy
|
||||
// a - dy
|
||||
GGML_API struct ggml_tensor * ggml_rope_back(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
int n_past,
|
||||
int n_dims,
|
||||
int mode);
|
||||
|
||||
// alibi position embedding
|
||||
// in-place, returns view(a)
|
||||
struct ggml_tensor * ggml_alibi(
|
||||
|
|
@ -740,13 +926,13 @@ extern "C" {
|
|||
GGML_API struct ggml_tensor * ggml_map_unary_f32(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
const ggml_unary_op_f32_t fun);
|
||||
ggml_unary_op_f32_t fun);
|
||||
|
||||
GGML_API struct ggml_tensor * ggml_map_binary_f32(
|
||||
struct ggml_context * ctx,
|
||||
struct ggml_tensor * a,
|
||||
struct ggml_tensor * b,
|
||||
const ggml_binary_op_f32_t fun);
|
||||
ggml_binary_op_f32_t fun);
|
||||
|
||||
//
|
||||
// automatic differentiation
|
||||
|
|
|
|||
16
llama.cpp
16
llama.cpp
|
|
@ -1128,8 +1128,8 @@ static bool llama_eval_internal(
|
|||
// self-attention
|
||||
{
|
||||
// compute Q and K and RoPE them
|
||||
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
||||
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
||||
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
||||
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
||||
ggml_set_name(Qcur, "Qcur");
|
||||
ggml_set_name(Kcur, "Kcur");
|
||||
|
||||
|
|
@ -1170,17 +1170,19 @@ static bool llama_eval_internal(
|
|||
struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head));
|
||||
ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_head)");
|
||||
|
||||
struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
|
||||
// KQ_scaled shape [n_past + N, N, n_head, 1]
|
||||
struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
|
||||
ggml_set_name(KQ_scaled, "KQ_scaled");
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
||||
ggml_set_name(KQ_masked, "KQ_masked");
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
|
||||
ggml_set_name(KQ_soft_max, "KQ_soft_max");
|
||||
|
||||
|
||||
// split cached V into n_head heads
|
||||
struct ggml_tensor * V =
|
||||
ggml_view_3d(ctx0, kv_self.v,
|
||||
|
|
@ -1281,7 +1283,7 @@ static bool llama_eval_internal(
|
|||
lctx.use_buf(ctx0, -1);
|
||||
|
||||
// logits -> probs
|
||||
//inpL = ggml_soft_max(ctx0, inpL);
|
||||
//inpL = ggml_soft_max_inplace(ctx0, inpL);
|
||||
|
||||
// run the computation
|
||||
ggml_build_forward_expand(&gf, inpL);
|
||||
|
|
@ -2375,7 +2377,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
|
|||
|
||||
if (scaling != 1.0f) {
|
||||
ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
|
||||
BA = ggml_scale(lora_ctx, BA, scale_tensor);
|
||||
BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
|
||||
}
|
||||
|
||||
ggml_tensor * r;
|
||||
|
|
|
|||
|
|
@ -10,3 +10,5 @@ llama_add_test(test-quantize-fns.cpp)
|
|||
llama_add_test(test-quantize-perf.cpp)
|
||||
llama_add_test(test-sampling.cpp)
|
||||
llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
|
||||
# llama_add_test(test-grad0.c) # SLOW
|
||||
# llama_add_test(test-opt.c) # SLOW
|
||||
|
|
|
|||
1131
tests/test-grad0.c
Normal file
1131
tests/test-grad0.c
Normal file
File diff suppressed because it is too large
Load diff
205
tests/test-opt.c
Normal file
205
tests/test-opt.c
Normal file
|
|
@ -0,0 +1,205 @@
|
|||
#include "ggml.h"
|
||||
|
||||
#include <math.h>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include <assert.h>
|
||||
|
||||
#define MAX_NARGS 2
|
||||
|
||||
|
||||
//
|
||||
// logging
|
||||
//
|
||||
#define GGML_DEBUG 0
|
||||
#if (GGML_DEBUG >= 1)
|
||||
#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
|
||||
#else
|
||||
#define GGML_PRINT_DEBUG(...)
|
||||
#endif
|
||||
|
||||
#if (GGML_DEBUG >= 5)
|
||||
#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
|
||||
#else
|
||||
#define GGML_PRINT_DEBUG_5(...)
|
||||
#endif
|
||||
|
||||
#if (GGML_DEBUG >= 10)
|
||||
#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
|
||||
#else
|
||||
#define GGML_PRINT_DEBUG_10(...)
|
||||
#endif
|
||||
|
||||
#define GGML_PRINT(...) printf(__VA_ARGS__)
|
||||
|
||||
|
||||
float frand() {
|
||||
return (float)rand()/(float)RAND_MAX;
|
||||
}
|
||||
|
||||
int irand(int n) {
|
||||
return rand()%n;
|
||||
}
|
||||
|
||||
void get_random_dims(int64_t * dims, int ndims) {
|
||||
dims[0] = dims[1] = dims[2] = dims[3] = 1;
|
||||
|
||||
for (int i = 0; i < ndims; i++) {
|
||||
dims[i] = 1 + irand(4);
|
||||
}
|
||||
}
|
||||
|
||||
void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
|
||||
dims[0] = dims[1] = dims[2] = dims[3] = 1;
|
||||
|
||||
for (int i = 0; i < ndims; i++) {
|
||||
dims[i] = min + irand(max-min);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
struct ggml_tensor * get_random_tensor(
|
||||
struct ggml_context * ctx0,
|
||||
int ndims,
|
||||
int64_t ne[],
|
||||
float fmin,
|
||||
float fmax) {
|
||||
struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
|
||||
|
||||
switch (ndims) {
|
||||
case 1:
|
||||
for (int i0 = 0; i0 < ne[0]; i0++) {
|
||||
((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
|
||||
}
|
||||
break;
|
||||
case 2:
|
||||
for (int i1 = 0; i1 < ne[1]; i1++) {
|
||||
for (int i0 = 0; i0 < ne[0]; i0++) {
|
||||
((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
|
||||
}
|
||||
}
|
||||
break;
|
||||
case 3:
|
||||
for (int i2 = 0; i2 < ne[2]; i2++) {
|
||||
for (int i1 = 0; i1 < ne[1]; i1++) {
|
||||
for (int i0 = 0; i0 < ne[0]; i0++) {
|
||||
((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
|
||||
}
|
||||
}
|
||||
}
|
||||
break;
|
||||
case 4:
|
||||
for (int i3 = 0; i3 < ne[3]; i3++) {
|
||||
for (int i2 = 0; i2 < ne[2]; i2++) {
|
||||
for (int i1 = 0; i1 < ne[1]; i1++) {
|
||||
for (int i0 = 0; i0 < ne[0]; i0++) {
|
||||
((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
break;
|
||||
default:
|
||||
assert(false);
|
||||
};
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
float get_element(const struct ggml_tensor * t, int idx) {
|
||||
return ((float *)t->data)[idx];
|
||||
}
|
||||
|
||||
void set_element(struct ggml_tensor * t, int idx, float value) {
|
||||
((float *)t->data)[idx] = value;
|
||||
}
|
||||
|
||||
int main(int argc, const char ** argv) {
|
||||
struct ggml_init_params params = {
|
||||
.mem_size = 1024*1024*1024,
|
||||
.mem_buffer = NULL,
|
||||
.no_alloc = false,
|
||||
};
|
||||
struct ggml_context * ctx = ggml_init(params);
|
||||
|
||||
int64_t ne1[4] = {4, 1024, 1, 1};
|
||||
int64_t ne2[4] = {4, 2048, 1, 1};;
|
||||
int64_t ne3[4] = {1024, 2048, 1, 1};
|
||||
|
||||
struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
|
||||
struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);
|
||||
ggml_set_param(ctx, a);
|
||||
ggml_set_param(ctx, b);
|
||||
|
||||
struct ggml_tensor * c = get_random_tensor(ctx, 2, ne3, -1, +1);
|
||||
|
||||
struct ggml_tensor * ab = ggml_mul_mat(ctx, a, b);
|
||||
struct ggml_tensor * d = ggml_sub(ctx, c, ab);
|
||||
struct ggml_tensor * e = ggml_sum(ctx, ggml_sqr(ctx, d));
|
||||
|
||||
|
||||
struct ggml_cgraph ge = ggml_build_forward(e);
|
||||
ggml_graph_reset (&ge);
|
||||
ggml_graph_compute(ctx, &ge);
|
||||
const float fe = ggml_get_f32_1d(e, 0);
|
||||
printf("%s: e = %.4f\n", __func__, fe);
|
||||
|
||||
struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM);
|
||||
|
||||
ggml_opt(ctx, opt_params, e);
|
||||
|
||||
ggml_graph_reset (&ge);
|
||||
ggml_graph_compute(ctx, &ge);
|
||||
const float fe_opt = ggml_get_f32_1d(e, 0);
|
||||
printf("%s: original e = %.4f\n", __func__, fe);
|
||||
printf("%s: optimized e = %.4f\n", __func__, fe_opt);
|
||||
|
||||
const bool success = (fe_opt <= fe);
|
||||
assert(success);
|
||||
|
||||
ggml_free(ctx);
|
||||
return success ? 0 : -1;
|
||||
}
|
||||
// int64_t ne1[4] = {4, 128, 1, 1};
|
||||
// int64_t ne2[4] = {4, 256, 1, 1};;
|
||||
// int64_t ne3[4] = {128, 256, 1, 1};
|
||||
// main: original e = 25890.9375
|
||||
// main: optimized e = 10094.7031
|
||||
|
||||
// int64_t ne1[4] = {8, 128, 1, 1};
|
||||
// int64_t ne2[4] = {8, 256, 1, 1};;
|
||||
// int64_t ne3[4] = {128, 256, 1, 1};
|
||||
// main: original e = 39429.5078
|
||||
// main: optimized e = 9275.8936
|
||||
|
||||
// int64_t ne1[4] = {16, 128, 1, 1};
|
||||
// int64_t ne2[4] = {16, 256, 1, 1};;
|
||||
// int64_t ne3[4] = {128, 256, 1, 1};
|
||||
// main: original e = 68371.1328
|
||||
// main: optimized e = 7854.4502
|
||||
|
||||
|
||||
// int64_t ne1[4] = {32, 128, 1, 1};
|
||||
// int64_t ne2[4] = {32, 256, 1, 1};;
|
||||
// int64_t ne3[4] = {128, 256, 1, 1};
|
||||
// main: original e = 126061.1953
|
||||
// main: optimized e = 5451.0166
|
||||
|
||||
// int64_t ne1[4] = {4, 1024, 1, 1};
|
||||
// int64_t ne2[4] = {4, 2048, 1, 1};;
|
||||
// int64_t ne3[4] = {1024, 2048, 1, 1};
|
||||
// main: original e = 1620817.8750
|
||||
// main: optimized e = 698387.6875
|
||||
|
||||
// another run on M1
|
||||
// int64_t ne1[4] = {4, 1024, 1, 1};
|
||||
// int64_t ne2[4] = {4, 2048, 1, 1};;
|
||||
// int64_t ne3[4] = {1024, 2048, 1, 1};
|
||||
// main: original e = 1629595.6250
|
||||
// main: optimized e = 698169.1250
|
||||
|
||||
// int64_t ne1[4] = {32, 1024, 1, 1};
|
||||
// int64_t ne2[4] = {32, 2048, 1, 1};;
|
||||
// int64_t ne3[4] = {1024, 2048, 1, 1};
|
||||
// main: original e = 8146770.5000
|
||||
// main: optimized e = 651119.1250
|
||||
Loading…
Reference in a new issue