vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama: add support for QRWKV6 model architecture (#11001)

Browse source 
llama: add support for QRWKV6 model architecture (#11001)

* WIP: Add support for RWKV6Qwen2

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* RWKV: Some graph simplification

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Add support for RWKV6Qwen2 with cpu and cuda GLA

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* RWKV6[QWEN2]: Concat lerp weights together to reduce cpu overhead

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix some typos

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* code format changes

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix wkv test & add gla test

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Fix cuda warning

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Update README.md

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Update ggml/src/ggml-cuda/gla.cu

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Fix fused lerp weights loading with RWKV6

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* better sanity check skipping for QRWKV6 in llama-quant

thanks @compilade

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
Co-authored-by: compilade <git@compilade.net>

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: compilade <git@compilade.net>
This commit is contained in:
Molly Sophia 2025-01-10 09:58:08 +08:00 • committed by GitHub
parent c6860cc734
commit ee7136c6d1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
23 changed files with 862 additions and 124 deletions
Download patch file Download diff file Expand all files Collapse all files

26
gguf-py/gguf/tensor_mapping.py
View file

@ -13,7 +13,7 @@ class TensorNameMap:
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf nemotron olmoe olmo2
"model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert nomic-bert
"language_model.embedding.word_embeddings", # persimmon
@ -464,34 +464,42 @@ class TensorNameMap:
MODEL_TENSOR.TIME_MIX_W1: (
"rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_W2: (
"rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_X: (
"rwkv.blocks.{bid}.attention.time_maa_x", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_x", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_K: (
"rwkv.blocks.{bid}.attention.time_maa_k", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_k", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_V: (
"rwkv.blocks.{bid}.attention.time_maa_v", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_v", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_R: (
"rwkv.blocks.{bid}.attention.time_maa_r", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_r", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_G: (
"rwkv.blocks.{bid}.attention.time_maa_g", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_g", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_W: (
"rwkv.blocks.{bid}.attention.time_maa_w", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_FIRST: (
@ -500,30 +508,37 @@ class TensorNameMap:
MODEL_TENSOR.TIME_MIX_DECAY: (
"rwkv.blocks.{bid}.attention.time_decay", # rwkv v6
"model.layers.{bid}.self_attn.time_decay", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W1: (
"rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv v6
"model.layers.{bid}.self_attn.time_decay_w1", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W2: (
"rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv v6
"model.layers.{bid}.self_attn.time_decay_w2", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_KEY: (
"rwkv.blocks.{bid}.attention.key", # rwkv
"rwkv.blocks.{bid}.attention.key", # rwkv
"model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_VALUE: (
"rwkv.blocks.{bid}.attention.value", # rwkv
"rwkv.blocks.{bid}.attention.value", # rwkv
"model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_RECEPTANCE: (
"rwkv.blocks.{bid}.attention.receptance", # rwkv
"model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_GATE: (
"rwkv.blocks.{bid}.attention.gate", # rwkv
"rwkv.blocks.{bid}.attention.gate", # rwkv
"model.layers.{bid}.self_attn.gate", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LN: (
@ -531,7 +546,8 @@ class TensorNameMap:
),
MODEL_TENSOR.TIME_MIX_OUTPUT: (
"rwkv.blocks.{bid}.attention.output", # rwkv
"rwkv.blocks.{bid}.attention.output", # rwkv
"model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2
),
MODEL_TENSOR.CHANNEL_MIX_LERP_K: (