xaedes
1151653b15
replace llama API functions to get model tensors by one function to get model tensor by name
...
LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
2023-08-16 21:36:40 +02:00
xaedes
39a2d15461
avoid stack overflow resulting from big ggml_cgraph
...
replace stack allocation and ggml_build_forward by ggml_new_graph in combination with ggml_build_forward_expand
2023-08-16 16:42:25 +02:00
xaedes
0ab2507ce5
fix names of lora tensors
2023-08-16 16:41:20 +02:00
xaedes
620275361d
add debug prints for training memory improvements
2023-08-16 16:23:21 +02:00
xaedes
be7e564b11
bug fixes to make finetune compile
...
automatic allocator does not work yet
2023-08-16 16:21:43 +02:00
xaedes
50b1e66200
remove const model and layer arguments in API functions for accessing model tensors
2023-08-16 16:21:02 +02:00
xaedes
28ee0c8583
first draft for LORA finetune training
2023-08-16 15:31:04 +02:00
xaedes
c0a372fd3d
add API functions to access remaining model parameters:
...
mult, head and rot
2023-08-16 15:30:31 +02:00
xaedes
9eb1ef8653
move and remove code
2023-08-15 14:03:02 +02:00
xaedes
5e059ace25
add stub example for finetuning, based on train-text-from-scratch
2023-08-15 13:54:28 +02:00
xaedes
316b0707f4
add API functions to access llama model tensors
2023-08-15 13:53:13 +02:00
xaedes
3b5515bbe0
reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator
...
with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory.
the computation results are the same
2023-08-14 22:09:36 +02:00
xaedes
56228461c8
fix memory "leak" in optimizers
...
each iteration a new cplan with new memory for work data was allocated.
now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data.
2023-08-14 21:12:02 +02:00
xaedes
3e6468b097
fix test when to create temporary backward graph
...
temporary backward graph is only necessary when using checkpointing
2023-08-14 20:57:18 +02:00
xaedes
098654c277
only use ggml_allocr_alloc when tensor has NULL data and is no view
2023-08-14 20:57:18 +02:00
xaedes
faf3e21eaf
add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly
2023-08-14 20:50:09 +02:00
xaedes
6e280b24dc
remove unused forward_batch function
2023-08-14 19:02:12 +02:00
xaedes
3794dceb7f
remove unused train params: mem_compute1_gb & mem_compute2_gb
...
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
2023-08-14 18:44:42 +02:00
xaedes
6f161c784b
remove trailing whitespace
2023-08-14 18:33:27 +02:00
xaedes
271e4d64b5
remove unused training parameters "use_scratch" and "use_unified"
2023-08-14 18:31:59 +02:00
xaedes
c954f41ca4
remove handwritten training functions
2023-08-14 18:30:50 +02:00
xaedes
fe788a1c7a
allocate graph on context using ggml_new_graph
2023-08-14 18:24:13 +02:00
xaedes
75baed230c
set names for tensors in unified train function for easier debugging
2023-08-14 18:17:14 +02:00
xaedes
3e99a8d653
format name of cloned tensors with " (clone)" suffix
2023-08-14 18:15:09 +02:00
xaedes
865c4cd3c1
integrate unified training function which may use memory allocator
...
the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing
2023-08-14 18:12:58 +02:00
xaedes
4ed096c6b0
add training options whether to use allocator and/or unified training function
2023-08-14 18:10:02 +02:00
xaedes
d6c5b03858
fix ASSERT to work with zero layers
2023-08-14 18:08:19 +02:00
xaedes
38f4438c32
make sure some tensors are not reallocated by inserting new temporary nodes depending on them:
...
output and parameter gradient tensors need to be available at the end of the graph execution
parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration
checkpoint tensors are allocated all together to reduce memory allocator fragmentation
afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs
2023-08-14 18:07:16 +02:00
xaedes
9716eb8ef0
fix variable name and add missing boolean negation
2023-08-14 17:59:19 +02:00
xaedes
5884b43a62
add input tensors as checkpoints
...
so that recursive tensor cloning of gradient checkpointing terminates on input tensors
2023-08-14 17:58:49 +02:00
xaedes
b2f1310196
swap arguments to commutative ops to be the same as in forward_batch_wo_cache_flash_attn
2023-08-14 17:57:13 +02:00
xaedes
5a11b75875
fix variable names
2023-08-14 17:55:51 +02:00
xaedes
345f516f7c
correctly clone view tensors by setting data pointers
...
without this the checkpointing would only work when being used together with memory allocator
2023-08-14 17:55:13 +02:00
xaedes
52c92c0a8c
terminate recursive tensor cloning when reaching tensor without src tensors
2023-08-14 17:53:36 +02:00
xaedes
0dd496c5e2
fix variable name and add missing type cast
2023-08-14 17:52:48 +02:00
xaedes
cfddc36be2
correctly clone reshape and permute operations by also cloning tensor->nb values
2023-08-14 17:52:15 +02:00
xaedes
d43741540b
don't use allocate hash_map on context
...
because the context has no_alloc=True when using memory allocator resulting in NULL data pointers
2023-08-14 17:51:20 +02:00
xaedes
fc826c8ea8
in train function replace add_inplace by regular add
...
because using add_inplace seems to result in different gradients
2023-08-14 17:49:22 +02:00
xaedes
2bf422eafd
add train function using automatic gradient checkpointing backward pass and allocator
2023-08-06 23:07:57 +02:00
xaedes
d43af4b543
Merge branch 'master' into pr-train-mem-usage-improvements
2023-08-06 17:30:17 +02:00
DannyDaemonic
86c3219895
console : fix issue related to Windows 11 PowerShell console mode persistence ( #2521 )
2023-08-06 09:49:34 +03:00
Keiichi Tabata
2e8265ae17
convert.py : add missing abstract methods for quantized data ( #2491 )
2023-08-06 09:34:05 +03:00
Johannes Gäßler
f514d1b306
CUDA: faster k-quant mul_mat_q kernels ( #2525 )
2023-08-05 18:20:44 +02:00
Jonas Wunderlich
332311234a
fix firefox autoscroll ( #2519 )
2023-08-04 22:16:11 +02:00
Cebtenzzre
182af739c4
server: regenerate completion.js.hpp ( #2515 )
2023-08-04 21:00:57 +02:00
Cebtenzzre
4329d1acb0
CUDA: use min compute capability of GPUs actually used ( #2506 )
2023-08-04 17:35:22 +02:00
Cebtenzzre
02f9d96a86
CUDA: check if event is NULL before cudaStreamWaitEvent ( #2505 )
...
Fixes #2503
2023-08-04 17:34:32 +02:00
DannyDaemonic
3498588e0f
Add --simple-io option for subprocesses and break out console.h and cpp ( #1558 )
2023-08-04 08:20:12 -07:00
Stephen Nichols
5f631c2679
Fixing race condition in server and partial stream handling in frontend. ( #2391 )
...
* Fixing race condition in server.cpp and partial stream handling in completion.js
* Reverting assert edits.
* Adding newline to eof
2023-08-04 13:37:24 +02:00
l3utterfly
415e99fec2
Stream save llama context data to file instead of allocating entire buffer upfront ( #2488 )
...
* added stream saving context data to file to avoid allocating unnecessary amounts of memory
* generalised copying state data to file or buffer
* added comments explaining how copy_state_data works
* fixed trailing whitespaces
* fixed save load state example
* updated save load state to use public function in llama.cpp
* - restored breakage of the llama_copy_state_data API
- moved new logic for copying llama state data to internal function
* fixed function declaration order
* restored save load state example
* fixed whitepace
* removed unused llama-util.h include
* Apply suggestions from code review
Co-authored-by: slaren <slarengh@gmail.com>
* Apply code review suggestions
Co-authored-by: slaren <slarengh@gmail.com>
---------
Co-authored-by: slaren <slarengh@gmail.com>
2023-08-04 13:29:52 +02:00
