llama : support RWKV v6 models #8980

MollySophia · 2024-08-11T02:09:47Z

This should fix #846.

Added:

ggml:

Added unary operation Exp
Added rwkv_wkv operation with CPU impl
Added rwkv_token_shift operation with CPU impl to handle multiple sequences in parallel(may not be necessary after llama : simplify Mamba with advanced batch splits #8526 is done)

llama.cpp:

rwkv_world tokenizer support (by @LaylBongers)
convert_hf_to_gguf.py support for converting RWKV v6 HF models
RWKV v6 graph building

TODO:

~~Do modifications after llama : simplify Mamba with advanced batch splits #8526 is ready accordingly~~ Done
~~Add CUDA or Metal implementation for rwkv_wkv operation~~ Maybe next PR

I have read the contributing guidelines
Self-reported review complexity:
- Medium

compilade

A few things I've noticed. I'll review this more deeply in the next days.

src/llama.cpp

convert_hf_to_gguf.py

ggml/src/ggml.c

convert_hf_to_gguf.py

src/llama-vocab.cpp

convert_hf_to_gguf.py

src/llama.cpp

MollySophia · 2024-08-23T02:17:42Z

Synchronized the changes and made it working again after #8526 being merged.
This PR should be ready for review again now :D
@compilade Could you take a look when convenient?

compilade

I'm impressed that ggml_rwkv_wkv only takes around 2% of the CPU time during inference of the 1.6B RWKV-v6 model (when measured with perf record --call-graph=lbr).

I have some styling comments, some suggestions, and I also found some problems.

src/llama.cpp

convert_hf_to_gguf.py

src/llama.cpp

MollySophia · 2024-08-25T04:31:44Z

I'm impressed that ggml_rwkv_wkv only takes around 2% of the CPU time during inference of the 1.6B RWKV-v6 model (when measured with perf record --call-graph=lbr).

I have some styling comments, some suggestions, and I also found some problems.

Indeed. I did consider writing a metal kernel for wkv, but it turned out that wkv kernels didn't eat much cpu time.
I've also tried modfying current rwkv_wkv impl with GGML_SIMD macros, but the speed was almost the same. (Clang already did optimizations like vectorization, so writing manually may not be that necessary)

src/llama.cpp

Signed-off-by: Molly Sophia <[email protected]>

Co-authored-by: compilade <[email protected]>

Signed-off-by: Molly Sophia <[email protected]>

…t tensors Signed-off-by: Molly Sophia <[email protected]>

Signed-off-by: Molly Sophia <[email protected]>

Currently att.key/receptance/value/gate/output, ffn.receptance/key/value, as well as head.weight Signed-off-by: Molly Sophia <[email protected]>

ggerganov · 2024-08-30T08:18:10Z

Lets look to merge soon. @MollySophia Which HF model do you recommend to run a few tests with this branch?

MollySophia · 2024-08-30T08:20:57Z

Lets look to merge soon. @MollySophia Which HF model do you recommend to run a few tests with this branch?

https://huggingface.co/RWKV/v6-Finch-1B6-HF should be enough for testing the functionalities.
https://huggingface.co/RWKV/v6-Finch-7B-HF/tree/main or the 3B one should be working too

ggerganov · 2024-08-30T10:22:32Z

I've updated the tokenizer to use a true for string search (7004323). With this change the time for tokenizing wiki.test dropped from 27s to 40ms on my Mac.

compilade

BTW What's next for this PR?

@MollySophia It looks ready for me, at least. Nice work!

There's some potential division by zero with hparams.rescale_every_n_layers which I think should be fixed before merging.

Improvements to ggml_rwkv_wkv (if relevant) can be done later in a follow-up PR, so I think this will be ready to merge.

ggml/src/ggml.c

src/llama.cpp

Co-authored-by: compilade <[email protected]>

Signed-off-by: Molly Sophia <[email protected]>

* convert_hf_to_gguf: Add support for RWKV v6 Signed-off-by: Molly Sophia <[email protected]> * Add RWKV tokenization * Fix build Signed-off-by: Molly Sophia <[email protected]> * Do not use special tokens when matching in RWKV tokenizer * Fix model loading * Add (broken) placeholder graph builder for RWKV * Add workaround for kv cache * Add logits conversion to rwkv5 * Add rwkv5 layer norms * Add time mix KVRG & correct merge mistake * Add remaining time mix parameters * Add time mix output loading * Add placeholder llm_build_time_mix * Fix build Signed-off-by: Molly Sophia <[email protected]> * Load more tensors for rwkv v6 Signed-off-by: Molly Sophia <[email protected]> * Fix rwkv tokenizer Signed-off-by: Molly Sophia <[email protected]> * ggml: Add unary operator Exp Signed-off-by: Molly Sophia <[email protected]> * RWKV v6 graph building Signed-off-by: Molly Sophia <[email protected]> * Add ``rescale_every_n_layers`` parameter Signed-off-by: Molly Sophia <[email protected]> * Add ``wkv.head_size`` key for RWKV so it doesn't reuse Mamba ssm parameters Signed-off-by: Molly Sophia <[email protected]> * Fix offloading layers to CUDA Signed-off-by: Molly Sophia <[email protected]> * Fix parallel inferencing for RWKV Signed-off-by: Molly Sophia <[email protected]> * Remove trailing whitespaces Signed-off-by: Molly Sophia <[email protected]> * build_rwkv: Avoid using inplace operations Signed-off-by: Molly Sophia <[email protected]> * convert_hf_to_gguf: rwkv: Avoid using ``eval`` Signed-off-by: Molly Sophia <[email protected]> * convert_hf_to_gguf: rwkv tokenizer: Don't escape sequences manually Signed-off-by: Molly Sophia <[email protected]> * Update convert_hf_to_gguf.py Co-authored-by: compilade <[email protected]> * ggml: Add backward computation for unary op ``exp`` Signed-off-by: Molly Sophia <[email protected]> * Update convert_hf_to_gguf.py Co-authored-by: compilade <[email protected]> * Update convert_hf_to_gguf.py Co-authored-by: compilade <[email protected]> * Use MODEL_ARCH.RWKV6 instead of MODEL_ARCH.RWKV Signed-off-by: Molly Sophia <[email protected]> * build_rwkv6: Simplify graph Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Detect model.type Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Fix tensor loading for 7B/14B models Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Fix group_norm assertion failure with Metal Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Clean up Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Add quantization tensor exclusion Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Use the new advanced batch splits Signed-off-by: Molly Sophia <[email protected]> * Update src/llama.cpp Co-authored-by: compilade <[email protected]> * llama: rwkv6: Use ``ggml_norm`` instead of ``ggml_group_norm`` Co-authored-by: compilade <[email protected]> * llama: rwkv6: Apply code style and misc changes Signed-off-by: Molly Sophia <[email protected]> * converter: Use class name ``Rwkv6Model`` Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Make use of key ``feed_forward_length`` Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Add kv ``time_mix_extra_dim`` and ``time_decay_extra_dim`` Signed-off-by: Molly Sophia <[email protected]> * converter: Match ``new_name`` instead of ``name`` for float32 explicit tensors Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Keep ``time_mix_w1/w2`` as F32 Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Remove unused nodes Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Apply code format changes Signed-off-by: Molly Sophia <[email protected]> * llama: rwkv6: Add lora for some supported tensors Currently att.key/receptance/value/gate/output, ffn.receptance/key/value, as well as head.weight Signed-off-by: Molly Sophia <[email protected]> * rwkv : speed-up tokenization using trie * minor : style + indentation * llama: rwkv6: Avoid division by zero Co-authored-by: compilade <[email protected]> * ggml: rwkv_wkv: Avoid copying the state Signed-off-by: Molly Sophia <[email protected]> --------- Signed-off-by: Molly Sophia <[email protected]> Co-authored-by: Layl Bongers <[email protected]> Co-authored-by: compilade <[email protected]> Co-authored-by: Georgi Gerganov <[email protected]>

github-actions bot added python python script changes ggml changes relating to the ggml tensor library for machine learning labels Aug 11, 2024

compilade self-requested a review August 11, 2024 02:30

MollySophia force-pushed the for-upstream branch from 5280749 to cf40fd3 Compare August 11, 2024 03:09

compilade reviewed Aug 11, 2024

View reviewed changes

src/llama.cpp Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

MollySophia force-pushed the for-upstream branch 2 times, most recently from 487fb6d to 9bf958f Compare August 11, 2024 04:11

Ronsor reviewed Aug 11, 2024

View reviewed changes

ggml/src/ggml.c Show resolved Hide resolved

Ronsor reviewed Aug 11, 2024

View reviewed changes

convert_hf_to_gguf.py Show resolved Hide resolved

Ronsor reviewed Aug 11, 2024

View reviewed changes

src/llama-vocab.cpp Show resolved Hide resolved

compilade reviewed Aug 12, 2024

View reviewed changes

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

MollySophia force-pushed the for-upstream branch from 6edbe81 to bc3e37d Compare August 12, 2024 01:13

compilade reviewed Aug 12, 2024

View reviewed changes

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

MollySophia force-pushed the for-upstream branch 3 times, most recently from ecf84ca to e7d35a3 Compare August 13, 2024 09:20

MollySophia commented Aug 13, 2024

View reviewed changes

src/llama.cpp Outdated Show resolved Hide resolved

rmusser01 mentioned this pull request Aug 17, 2024

Feature Request: Support Codestral Mamba #8519

Open

compilade mentioned this pull request Aug 19, 2024

llama : simplify Mamba with advanced batch splits #8526

Merged

10 tasks

MollySophia force-pushed the for-upstream branch 2 times, most recently from d7e71a5 to c3564d8 Compare August 23, 2024 02:14

compilade reviewed Aug 25, 2024

View reviewed changes

MollySophia force-pushed the for-upstream branch 2 times, most recently from 8e2e9aa to a8db247 Compare August 25, 2024 09:36

compilade reviewed Aug 26, 2024

View reviewed changes

src/llama.cpp Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

MollySophia mentioned this pull request Aug 26, 2024

Add support for loading RWKV v6 GGUF files RWKV/rwkv.cpp#180

Open

MollySophia and others added 2 commits August 28, 2024 10:20

convert_hf_to_gguf: Add support for RWKV v6

8d2eca3

Signed-off-by: Molly Sophia <[email protected]>

Add RWKV tokenization

dc0767f

MollySophia and others added 11 commits August 28, 2024 10:22

llama: rwkv6: Use the new advanced batch splits

f5d955d

Signed-off-by: Molly Sophia <[email protected]>

Update src/llama.cpp

57decb4

Co-authored-by: compilade <[email protected]>

llama: rwkv6: Use ggml_norm instead of ggml_group_norm

e94778a

Co-authored-by: compilade <[email protected]>

llama: rwkv6: Apply code style and misc changes

7756afd

Signed-off-by: Molly Sophia <[email protected]>

converter: Use class name Rwkv6Model

87a2901

Signed-off-by: Molly Sophia <[email protected]>

llama: rwkv6: Make use of key feed_forward_length

c414a24

Signed-off-by: Molly Sophia <[email protected]>

llama: rwkv6: Add kv time_mix_extra_dim and time_decay_extra_dim

6d69fd7

Signed-off-by: Molly Sophia <[email protected]>

converter: Match new_name instead of name for float32 explici…

601b592

…t tensors Signed-off-by: Molly Sophia <[email protected]>

llama: rwkv6: Keep time_mix_w1/w2 as F32

e0ea511

Signed-off-by: Molly Sophia <[email protected]>

llama: rwkv6: Remove unused nodes

5f00c52

Signed-off-by: Molly Sophia <[email protected]>

llama: rwkv6: Apply code format changes

7444046

Signed-off-by: Molly Sophia <[email protected]>

MollySophia force-pushed the for-upstream branch from a1429c2 to 7444046 Compare August 28, 2024 02:46

llama: rwkv6: Add lora for some supported tensors

7f2ef56

Currently att.key/receptance/value/gate/output, ffn.receptance/key/value, as well as head.weight Signed-off-by: Molly Sophia <[email protected]>

rwkv : speed-up tokenization using trie

7004323

minor : style + indentation

59dc2e7

ggerganov approved these changes Aug 30, 2024

View reviewed changes

ggerganov requested a review from compilade August 30, 2024 10:31

compilade approved these changes Aug 30, 2024

View reviewed changes

ggml/src/ggml.c Show resolved Hide resolved

ggml/src/ggml.c Outdated Show resolved Hide resolved

src/llama.cpp Outdated Show resolved Hide resolved

MollySophia and others added 2 commits August 31, 2024 11:59

llama: rwkv6: Avoid division by zero

5175375

Co-authored-by: compilade <[email protected]>

ggml: rwkv_wkv: Avoid copying the state

846358d

Signed-off-by: Molly Sophia <[email protected]>

ggerganov merged commit 8f1d81a into ggerganov:master Sep 1, 2024
54 checks passed

wszgrcy mentioned this pull request Sep 2, 2024

Add support for RWKV ollama/ollama#1612

Open

ggerganov mentioned this pull request Sep 3, 2024

changelog : libllama API #9289

Open

AndriyMulyar mentioned this pull request Sep 3, 2024

[Feature] Add support for RWKV nomic-ai/gpt4all#2933

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

llama : support RWKV v6 models #8980

llama : support RWKV v6 models #8980

MollySophia commented Aug 11, 2024 •

edited

Loading

compilade left a comment

MollySophia commented Aug 23, 2024 •

edited

Loading

compilade left a comment •

edited

Loading

MollySophia commented Aug 25, 2024

ggerganov commented Aug 30, 2024

MollySophia commented Aug 30, 2024

ggerganov commented Aug 30, 2024

compilade left a comment •

edited

Loading

llama : support RWKV v6 models #8980

llama : support RWKV v6 models #8980

Conversation

MollySophia commented Aug 11, 2024 • edited Loading

Added:

ggml:

llama.cpp:

TODO:

compilade left a comment

Choose a reason for hiding this comment

MollySophia commented Aug 23, 2024 • edited Loading

compilade left a comment • edited Loading

Choose a reason for hiding this comment

MollySophia commented Aug 25, 2024

ggerganov commented Aug 30, 2024

MollySophia commented Aug 30, 2024

ggerganov commented Aug 30, 2024

compilade left a comment • edited Loading

Choose a reason for hiding this comment

MollySophia commented Aug 11, 2024 •

edited

Loading

MollySophia commented Aug 23, 2024 •

edited

Loading

compilade left a comment •

edited

Loading

compilade left a comment •

edited

Loading