ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama: Add support for RWKV v7 architecture (#12412)

Browse source 
* ggml: Add op l2_norm

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

* ggml: Add op rwkv_wkv7

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

* llama: Add support for RWKV7 and ARWKV7 models

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

* llama: fix inference with RWKV6Qwen2

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

* llama: add more (a)rwkv7 variants in size

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

* Apply code-format changes

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

* fix MUSA build

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

* llama: fix shape error with rwkv using llama-parallel

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

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
This commit is contained in:
Molly Sophia 2025-03-18 07:27:50 +08:00 • committed by GitHub
parent 60c902926c
commit 7dfad387e3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
35 changed files with 2948 additions and 438 deletions
Download patch file Download diff file Expand all files Collapse all files

131
gguf-py/gguf/tensor_mapping.py
View file

@ -27,7 +27,8 @@ class TensorNameMap:
"embedding.word_embeddings", # chatglm
"transformer.token_embeddings", # openelm
"shared", # t5
"rwkv.embeddings", # rwkv
"rwkv.embeddings", # rwkv6
"model.embeddings", # rwkv7
),
# Token type embeddings
@ -42,6 +43,9 @@ class TensorNameMap:
"emb_ln", # nomic-bert
"transformer.norm", # openelm
"rwkv.blocks.0.pre_ln", # rwkv
"rwkv.blocks.0.pre_ln", # rwkv6
"model.pre_ln", # rwkv7
"model.layers.0.pre_norm", # rwkv7
"backbone.norm", # wavtokenizer
),
@ -81,7 +85,8 @@ class TensorNameMap:
"encoder.final_layernorm", # chatglm
"transformer.norm", # openelm
"model.norm", # nemotron
"rwkv.ln_out", # rwkv
"rwkv.ln_out", # rwkv6
"model.ln_out", # rwkv7
"backbone.final_layer_norm", # wavtokenizer
),
@ -122,14 +127,16 @@ class TensorNameMap:
"transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx
"encoder.layers.{bid}.input_layernorm", # chatglm
"transformer.layers.{bid}.attn_norm", # openelm
"rwkv.blocks.{bid}.ln1", # rwkv
"rwkv.blocks.{bid}.ln1", # rwkv6
"model.layers.{bid}.ln1", # rwkv7
),
# Attention norm 2
MODEL_TENSOR.ATTN_NORM_2: (
"transformer.h.{bid}.ln_attn", # falcon40b
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
"rwkv.blocks.{bid}.ln2", # rwkv
"rwkv.blocks.{bid}.ln2", # rwkv6
"model.layers.{bid}.ln2", # rwkv7
),
# Attention query-key-value
@ -462,112 +469,174 @@ class TensorNameMap:
"backbone.layers.{bid}.mixer.out_proj",
),
MODEL_TENSOR.TIME_MIX_W0: (
"model.layers.{bid}.attention.w0", # rwkv7
),
MODEL_TENSOR.TIME_MIX_W1: (
"rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2
"rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv6
"model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2
"model.layers.{bid}.attention.w1", # rwkv7
),
MODEL_TENSOR.TIME_MIX_W2: (
"rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2
"rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv6
"model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2
"model.layers.{bid}.attention.w2", # rwkv7
),
MODEL_TENSOR.TIME_MIX_A0: (
"model.layers.{bid}.attention.a0", # rwkv7
),
MODEL_TENSOR.TIME_MIX_A1: (
"model.layers.{bid}.attention.a1", # rwkv7
),
MODEL_TENSOR.TIME_MIX_A2: (
"model.layers.{bid}.attention.a2", # rwkv7
),
MODEL_TENSOR.TIME_MIX_V0: (
"model.layers.{bid}.attention.v0", # rwkv7
),
MODEL_TENSOR.TIME_MIX_V1: (
"model.layers.{bid}.attention.v1", # rwkv7
),
MODEL_TENSOR.TIME_MIX_V2: (
"model.layers.{bid}.attention.v2", # rwkv7
),
MODEL_TENSOR.TIME_MIX_G1: (
"model.layers.{bid}.attention.g1", # rwkv7
),
MODEL_TENSOR.TIME_MIX_G2: (
"model.layers.{bid}.attention.g2", # rwkv7
),
MODEL_TENSOR.TIME_MIX_K_K: (
"model.layers.{bid}.attention.k_k", # rwkv7
),
MODEL_TENSOR.TIME_MIX_K_A: (
"model.layers.{bid}.attention.k_a", # rwkv7
),
MODEL_TENSOR.TIME_MIX_R_K: (
"model.layers.{bid}.attention.r_k", # rwkv7
),
MODEL_TENSOR.TIME_MIX_LERP_X: (
"rwkv.blocks.{bid}.attention.time_maa_x", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_x", # rwkv6
"model.layers.{bid}.self_attn.time_maa_x", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_K: (
"rwkv.blocks.{bid}.attention.time_maa_k", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_k", # rwkv6
"model.layers.{bid}.self_attn.time_maa_k", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_V: (
"rwkv.blocks.{bid}.attention.time_maa_v", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_v", # rwkv6
"model.layers.{bid}.self_attn.time_maa_v", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_R: (
"rwkv.blocks.{bid}.attention.time_maa_r", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_r", # rwkv6
"model.layers.{bid}.self_attn.time_maa_r", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_G: (
"rwkv.blocks.{bid}.attention.time_maa_g", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_g", # rwkv6
"model.layers.{bid}.self_attn.time_maa_g", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_W: (
"rwkv.blocks.{bid}.attention.time_maa_w", # rwkv v6
"rwkv.blocks.{bid}.attention.time_maa_w", # rwkv6
"model.layers.{bid}.self_attn.time_maa_w", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_FIRST: (
"rwkv.blocks.{bid}.attention.time_faaaa", # rwkv v6
"rwkv.blocks.{bid}.attention.time_faaaa", # rwkv6
),
MODEL_TENSOR.TIME_MIX_DECAY: (
"rwkv.blocks.{bid}.attention.time_decay", # rwkv v6
"rwkv.blocks.{bid}.attention.time_decay", # rwkv6
"model.layers.{bid}.self_attn.time_decay", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W1: (
"rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv v6
"rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv6
"model.layers.{bid}.self_attn.time_decay_w1", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W2: (
"rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv v6
"rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv6
"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", # rwkv6
"model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2
"model.layers.{bid}.attention.key", # rwkv7
"model.layers.{bid}.attention.k_proj", # rwkv7
),
MODEL_TENSOR.TIME_MIX_VALUE: (
"rwkv.blocks.{bid}.attention.value", # rwkv
"rwkv.blocks.{bid}.attention.value", # rwkv6
"model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2
"model.layers.{bid}.attention.value", # rwkv7
"model.layers.{bid}.attention.v_proj", # rwkv7
),
MODEL_TENSOR.TIME_MIX_RECEPTANCE: (
"rwkv.blocks.{bid}.attention.receptance", # rwkv
"model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2
"rwkv.blocks.{bid}.attention.receptance", # rwkv6
"model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2
"model.layers.{bid}.attention.receptance", # rwkv7
"model.layers.{bid}.attention.r_proj", # rwkv7
),
MODEL_TENSOR.TIME_MIX_GATE: (
"rwkv.blocks.{bid}.attention.gate", # rwkv
"model.layers.{bid}.self_attn.gate", # rwkv6qwen2
"rwkv.blocks.{bid}.attention.gate", # rwkv6
"model.layers.{bid}.self_attn.gate", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LN: (
"rwkv.blocks.{bid}.attention.ln_x", # rwkv
"rwkv.blocks.{bid}.attention.ln_x", # rwkv6
"model.layers.{bid}.attention.ln_x" # rwkv7
),
MODEL_TENSOR.TIME_MIX_OUTPUT: (
"rwkv.blocks.{bid}.attention.output", # rwkv
"rwkv.blocks.{bid}.attention.output", # rwkv6
"model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2
"model.layers.{bid}.attention.output", # rwkv7
"model.layers.{bid}.attention.o_proj", # rwkv7
),
MODEL_TENSOR.CHANNEL_MIX_LERP_K: (
"rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv v6
"rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv6
"model.layers.{bid}.feed_forward.x_k", # rwkv7
),
MODEL_TENSOR.CHANNEL_MIX_LERP_R: (
"rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv v6
"rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv6
),
MODEL_TENSOR.CHANNEL_MIX_KEY: (
"rwkv.blocks.{bid}.feed_forward.key", # rwkv
"rwkv.blocks.{bid}.feed_forward.key", # rwkv6
"model.layers.{bid}.feed_forward.key", # rwkv7
),
MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: (
"rwkv.blocks.{bid}.feed_forward.receptance", # rwkv
"rwkv.blocks.{bid}.feed_forward.receptance", # rwkv6
),
MODEL_TENSOR.CHANNEL_MIX_VALUE: (
"rwkv.blocks.{bid}.feed_forward.value", # rwkv
"rwkv.blocks.{bid}.feed_forward.value", # rwkv6
"model.layers.{bid}.feed_forward.value", # rwkv7
),
MODEL_TENSOR.ATTN_Q_A: (