Model: Qwen3 Next #16095
Model: Qwen3 Next #16095
Conversation
github-actions
bot
added
python
|
I'll try to get into it in more detail soon, but here are a few general thoughts after quickly skimming the PR:
|
interesting, maybe we can learn together |
ggml/src/ggml.c
Outdated
| if (use_qk_l2norm) { | ||
| q_norm = ggml_l2_norm(ctx, q, 1e-6f); | ||
| k_norm = ggml_l2_norm(ctx, k, 1e-6f); | ||
| } | ||
|
|
||
| // Apply scaling to query | ||
| q_norm = ggml_scale(ctx, q_norm, scale); | ||
|
|
||
| // Apply sigmoid to beta for gating | ||
| struct ggml_tensor * beta_sigmoid = ggml_sigmoid(ctx, beta); | ||
| struct ggml_tensor * mixed_qkv = ggml_concat(ctx, q_norm, k_norm, 1); | ||
| mixed_qkv = ggml_concat(ctx, mixed_qkv, v, 1); | ||
|
|
||
| u_int32_t dim = (S_v * H_v) + 2 * (H_k * S_k); | ||
|
|
||
| mixed_qkv = ggml_reshape_3d(ctx, mixed_qkv, 1, dim, n_tokens); | ||
| struct ggml_tensor * mixed_qkv_padded = ggml_pad(ctx, mixed_qkv, 3, 0, 0, 0); |
There was a problem hiding this comment.
This part of code has namy magic number and configs (like l2norm, sigmoid, silu). It will be a headache if a future model reuse this delta net idea with some tweaks. It's better to just move al this part to ggml-model and the make ggml_delta_net being a thin wrapper around GGML_OP_DELTA_NET, like all other ops.
| int64_t ne3) { | ||
| GGML_ASSERT(ggml_is_contiguous(a)); | ||
| GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); | ||
| GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); |
There was a problem hiding this comment.
| GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); | |
| GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); |
ggml/src/ggml.c
Outdated
| q_broadcast = ggml_repeat_4d(ctx, q_broadcast, S_k, repeat_factor, H_k, n_tokens); | ||
| k_broadcast = ggml_repeat_4d(ctx, k_broadcast, S_k, repeat_factor, H_k, n_tokens); |
There was a problem hiding this comment.
maybe repeat_factor can be a param for GGML_OP_DELTA_NET, so it can internally do the broadcast without using extra memory
ggml/src/ggml.c
Outdated
| k_conv = ggml_permute(ctx, k_conv, 0, 2, 1, 3); | ||
| v_conv = ggml_permute(ctx, v_conv, 0, 2, 1, 3); | ||
|
|
||
| q_conv = ggml_reshape_3d(ctx, ggml_cont(ctx, q_conv), S_k * H_k, 1, n_tokens); |
There was a problem hiding this comment.
ggml_cont_3d is the combination of reshape and cont
ggml/src/ggml-cpu/ggml-cpu.c
Outdated
| } | ||
|
|
||
| // Apply sigmoid to beta | ||
| float * beta_sigmoid = (float *)alloca(n_tokens * sizeof(float)); |
There was a problem hiding this comment.
using working data (params->wdata) can be a better choice
ggml/src/ggml-cpu/ggml-cpu.c
Outdated
| // Apply sigmoid to beta | ||
| float * beta_sigmoid = (float *)alloca(n_tokens * sizeof(float)); | ||
| for (int64_t t = 0; t < n_tokens; ++t) { | ||
| beta_sigmoid[t] = 1.0f / (1.0f + expf(-beta_ptr[t * nb42 / sizeof(float)])); |
There was a problem hiding this comment.
isn't beta already be sigmoid-ed before passing to this op? you're doing sigmoid 2nd time here IIUC
ggml/src/ggml-cpu/ggml-cpu.c
Outdated
|
|
||
| // ggml_compute_forward_delta_net | ||
|
|
||
| static void ggml_compute_forward_delta_net( |
There was a problem hiding this comment.
I feel like this op can be implemented using other ggml ops like mul, mul_mat, sum. Which part of the calculation do you think that can't be constructed using existing ops?
Running #0 __syscall_cancel_arch () at ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S:56
56 in ../sysdeps/unix/sysv/linux/x86_64/syscall_cancel.S
#1 0x000070552b29eb63 in __internal_syscall_cancel (a1=<optimized out>, a2=<optimized out>, a3=<optimized out>, a4=<optimized out>, a5=0, a6=0, nr=61) at ./nptl/cancellation.c:49
warning: 49 ./nptl/cancellation.c: No such file or directory
#2 __syscall_cancel (a1=<optimized out>, a2=<optimized out>, a3=<optimized out>, a4=<optimized out>, a5=a5@entry=0, a6=a6@entry=0, nr=61) at ./nptl/cancellation.c:75
75 in ./nptl/cancellation.c
#3 0x000070552b31afdf in __GI___wait4 (pid=<optimized out>, stat_loc=<optimized out>, options=<optimized out>, usage=<optimized out>) at ../sysdeps/unix/sysv/linux/wait4.c:30
warning: 30 ../sysdeps/unix/sysv/linux/wait4.c: No such file or directory
#4 0x000070552bb45c31 in ggml_print_backtrace () at /devel/tools/llama.cpp/ggml/src/ggml.c:196
warning: Source file is more recent than executable.
196 waitpid(child_pid, NULL, 0);
#5 0x000070552bb45de5 in ggml_abort (file=0x70552bbcdac8 "/devel/tools/llama.cpp/ggml/src/ggml-backend.cpp", line=189, fmt=0x70552bbcd8af "GGML_ASSERT(%s) failed") at /devel/tools/llama.cpp/ggml/src/ggml.c:230
230 ggml_print_backtrace();
#6 0x000070552bb6091e in ggml_backend_buffer_get_type (buffer=0x0) at /devel/tools/llama.cpp/ggml/src/ggml-backend.cpp:189
189 GGML_ASSERT(buffer);
#7 0x000070552bb6080e in ggml_backend_buffer_is_host (buffer=0x0) at /devel/tools/llama.cpp/ggml/src/ggml-backend.cpp:170
170 return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
#8 0x000070552c07a114 in llm_graph_input_rs::set_input (this=0x5f11bdf6aea0, ubatch=0x5f11be011300) at /devel/tools/llama.cpp/src/llama-graph.cpp:241
241 GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
#9 0x000070552c07b03c in llm_graph_input_mem_hybrid::set_input (this=0x5f11bdf6aee0, ubatch=0x5f11be011300) at /devel/tools/llama.cpp/src/llama-graph.cpp:437
437 inp_rs->set_input(ubatch);
#10 0x000070552c07b549 in llm_graph_result::set_inputs (this=0x5f11be01ddf0, ubatch=0x5f11be011300) at /devel/tools/llama.cpp/src/llama-graph.cpp:480
480 input->set_input(ubatch);
#11 0x000070552c01ddb3 in llama_context::process_ubatch (this=0x5f11c05b5b50, ubatch=..., gtype=LLM_GRAPH_TYPE_DECODER, mctx=0x5f11be00ff00, ret=@0x7fff74d22ea4: 538976288) at /devel/tools/llama.cpp/src/llama-context.cpp:779
779 res->set_inputs(&ubatch);
#12 0x000070552c01f367 in llama_context::decode (this=0x5f11c05b5b50, batch_inp=...) at /devel/tools/llama.cpp/src/llama-context.cpp:1088
1088 const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
#13 0x000070552c025e49 in llama_decode (ctx=0x5f11c05b5b50, batch=...) at /devel/tools/llama.cpp/src/llama-context.cpp:2726
2726 const int ret = ctx->decode(batch);
#14 0x00005f11a2021559 in common_init_from_params (params=...) at /devel/tools/llama.cpp/common/common.cpp:1066
1066 llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
#15 0x00005f11a1e4a3c0 in main (argc=7, argv=0x7fff74d25968) at /devel/tools/llama.cpp/tools/main/main.cpp:140
140 common_init_result llama_init = common_init_from_params(params);I'll try to merge the op into the ggml_delta_net function call as @ngxson suggested. |
The backend buffer is NULL. |
The model doesn't seem to have any recurrence layers. This makes the set input fails due to input node not being present in cgraph.
Hmm I think I said the reverse: not to merge it but make the op simple
This is the more important question: should we try to implement it using existing ops, or add a new op and spend even more time to optimize it cross all backends? |
|
Now this is an error I haven't expected to encounter:
|
How do I allocate the memory for the linear layers then? I seem to have misunderstood how |
|
@pwilkin any chance to buy you a coffee?(Paterson etc.) so community able to donate for your efforts. Thank you! |
Added a buymeacoffee link to my profile (do consider first funding the Llama.cpp project itself, though!) |
I send a coffee also. |
Probably there are too many nodes on cgraph, try increasing the limit via |
src/llama-model.cpp
Outdated
| Qcur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, Qcur), n_embd_head, hparams.n_head(il), n_tokens); | ||
| Kcur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, Kcur), n_embd_head, hparams.n_head_kv(il), n_tokens); | ||
| Vcur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, Vcur), n_embd_head, hparams.n_head_kv(il), n_tokens); |
There was a problem hiding this comment.
these ggml_cont can be removed if Q/gate are separated. ggml_cont is not recommended when dealing with big tensors
There was a problem hiding this comment.
Actually none of these need ggml_cont, Q is 3D already, Q/K are RoPEd so can be views and V can also be a 3D view now.
Edit: sorry, not quite true about V, only if QKV is fused, the weird gate fuse threw me off. Nevertheless, K/V are already contiguous at this point.
There was a problem hiding this comment.
the problem is that Q is non-contiguous and ggml_rope(_ext) does not work very well with non-cont tensors, it's still buggy on certain backends
There was a problem hiding this comment.
the problem is that Q is non-contiguous and
ggml_rope(_ext)does not work very well with non-cont tensors, it's still buggy on certain backends
Are you sure? AFAIK those issues are fixed.
Edit: Also, if there still are issues they will never get fixed if we work around them. :)
There was a problem hiding this comment.
the problem is that Q is non-contiguous and
ggml_rope(_ext)does not work very well with non-cont tensors, it's still buggy on certain backends
I think all of these cases are fixed now.
There was a problem hiding this comment.
This was an impl of 2D rope that relies on ggml_view: https://github.com/ngxson/ggml-easy/blob/f56e5e499b1f21a4aae73010e9d9582840428457/demo/2d-rope.cpp
It works on CPU and Metal, but doesn't work on CUDA/Vulkan. Couldn't tested on other backends, but feel free to make a PR to address this issue.
There was a problem hiding this comment.
Does it still fail? I think these PRs should have addressed the problem:
There was a problem hiding this comment.
yes that seems to work. sorry @pwilkin you will need to manually revert the change where I split Q/gate. the tensor shape for Q will be:
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0);
src/llama-model.cpp
Outdated
| layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), { hparams.ssm_dt_rank }, 0); | ||
| layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_projection_size }, 0); | ||
| layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); | ||
| layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { n_ff, n_embd }, 0); |
There was a problem hiding this comment.
Shape of LLM_TENSOR_ATTN_Q and LLM_TENSOR_SSM_OUT should not contain n_ff
|
^ proposed fix for the 3 comments above: 46110e0 |
src/llama-model.cpp
Outdated
|
|
||
| GGML_ASSERT(ggml_nelements(beta) + ggml_nelements(alpha) == ggml_nelements(mixed_ba)); | ||
|
|
||
| // Softplus would be nice... |
There was a problem hiding this comment.
Better version of this code block: f643b95
src/llama-model.cpp
Outdated
| ggml_tensor * conv_bias = nullptr; // Add if your model has conv bias | ||
|
|
||
| // Get recurrent states (conv_states not needed as it's handled internally by ggml_delta_net) | ||
| ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); |
There was a problem hiding this comment.
mctx_cur->get_r_l is never used which make its buffer not initialized, and thus set_input fails
ggml/src/ggml-cpu/ggml-cpu.c
Outdated
| } | ||
|
|
||
| // Copy final state to new_state | ||
| memcpy(new_state_ptr, state_ptr, S * S * sizeof(float)); |
There was a problem hiding this comment.
state_ptr should be immutable, aka being const. you must store the result into dst, which will be ggml_cpy back into ssm_states_all
doing memcpy like this only works if all tensors are on CPU. if one on CPU and the other on GPU, this will crash.
There was a problem hiding this comment.
But ultimately I don't think we need a dedicated op for this. I think this is more or less the same thing with llm_graph_context_mamba::build_rs()
|
@ngxson Thanks, I got an LLM to rewrite the internal delta into tensor logic. After a day of manually fixing that crap, I think I understand it enough to rewrite it myself ;) |
|
Honestly I would prefer taking time to understand the mamba/ssm implementation then writing the code manually. Code written by LLM are mostly attempts for 1-to-1 translation from pytorch --> GGML which looks quite confusing |
Yeah, for me getting a rough outline then going over it manually is the best way to learn :) I tried the "one-to-one" approach and ended up with a graph that wouldn't fit in 16 GB of RAM for a 500M model... |
|
Aight, I cleaned up the main graph calculation, now I have to figure out how to include |
|
The parallelized version mitigates this a bit, but still I wouldn't recommend using anything above bs=128:
|
|
With "poor man's chunking" aka batch size = 64 it performs pretty OK on pp1024:
|
|
Yeah, it's starting to make sense. I think we can easily add intra-block (i.e. intra-ubatch) chunking in the future with the current approach. For now, don't think we should worry about that. It's still not clear to me why the implementations (i.e. the transformers from earlier and now the Kimi paper) use the unrolled version of |
# Conflicts: # convert_hf_to_gguf.py # gguf-py/gguf/constants.py # src/llama-arch.cpp # src/llama-arch.h # src/llama-model.cpp
Don't think I didn't try :) the current version of the graph for the real model has over 9000 nodes... The unrolled version, when I tried that, wouldn't fit in 16 GB of RAM for a 200M mock model. Anyways, I really think we should do something (
I thought that too, but our approach got validated by Kimi Linear's implementation, which does in fact use triangular solve for its chunked implementation: def chunk_kda_fwd(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
g: torch.Tensor,
beta: torch.Tensor,
scale: float,
initial_state: torch.Tensor,
output_final_state: bool,
cu_seqlens: torch.LongTensor | None = None,
):
chunk_size = 64
g = chunk_local_cumsum(g, chunk_size=chunk_size, cu_seqlens=cu_seqlens)
# the intra Aqk is kept in fp32
# the computation has very marginal effect on the entire throughput
A, Aqk = chunk_kda_scaled_dot_kkt_fwd(
q=q,
k=k,
gk=g,
beta=beta,
scale=scale,
cu_seqlens=cu_seqlens,
output_dtype=torch.float32,
)
A = solve_tril(A=A, cu_seqlens=cu_seqlens, output_dtype=k.dtype)
(...) |
|
Oh yeah, I've merged it with current master, so technically it should be ready for re-review :) the optimized CPU version of |
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This comment was marked as off-topic.
This comment was marked as off-topic.
|
Not sure if it is related to this PR. When evaulating Qwen3-Next with Vulkan backend on iGPU, the model was loaded to both CPU and Vulkan, which doubles the RAM. load_tensors: loading model tensors, this can take a while... (mmap = true) See logs
PS E:\LLM\qwen3_next_llama> .\build\bin\Release\llama-cli.exe -m E:\LLM\Qwen3-Next-80B-A3B-Instruct-IQ1_S_M.gguf
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon 780M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
build: 7337 (61667c3) with MSVC 19.44.35217.0 for x64
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon 780M Graphics) (unknown id) - 46478 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 807 tensors from E:\LLM\Qwen3-Next-80B-A3B-Instruct-IQ1_S_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = qwen3next
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Qwen3 Next 80B A3B Instruct
llama_model_loader: - kv 3: general.finetune str = Instruct
llama_model_loader: - kv 4: general.basename str = Qwen3-Next
llama_model_loader: - kv 5: general.size_label str = 80B-A3B
llama_model_loader: - kv 6: general.license str = apache-2.0
llama_model_loader: - kv 7: general.license.link str = https://huggingface.co/Qwen/Qwen3-Nex...
llama_model_loader: - kv 8: general.tags arr[str,1] = ["text-generation"]
llama_model_loader: - kv 9: qwen3next.block_count u32 = 48
llama_model_loader: - kv 10: qwen3next.context_length u32 = 262144
llama_model_loader: - kv 11: qwen3next.embedding_length u32 = 2048
llama_model_loader: - kv 12: qwen3next.feed_forward_length u32 = 5120
llama_model_loader: - kv 13: qwen3next.attention.head_count u32 = 16
llama_model_loader: - kv 14: qwen3next.attention.head_count_kv u32 = 2
llama_model_loader: - kv 15: qwen3next.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 16: qwen3next.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 17: qwen3next.expert_used_count u32 = 10
llama_model_loader: - kv 18: qwen3next.attention.key_length u32 = 256
llama_model_loader: - kv 19: qwen3next.attention.value_length u32 = 256
llama_model_loader: - kv 20: qwen3next.expert_count u32 = 512
llama_model_loader: - kv 21: qwen3next.expert_feed_forward_length u32 = 512
llama_model_loader: - kv 22: qwen3next.expert_shared_feed_forward_length u32 = 512
llama_model_loader: - kv 23: qwen3next.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 24: qwen3next.ssm.state_size u32 = 128
llama_model_loader: - kv 25: qwen3next.ssm.group_count u32 = 16
llama_model_loader: - kv 26: qwen3next.ssm.time_step_rank u32 = 32
llama_model_loader: - kv 27: qwen3next.ssm.inner_size u32 = 4096
llama_model_loader: - kv 28: qwen3next.rope.dimension_count u32 = 64
llama_model_loader: - kv 29: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 30: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 31: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 32: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 33: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 34: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 35: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 36: tokenizer.ggml.bos_token_id u32 = 151643
llama_model_loader: - kv 37: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 38: tokenizer.chat_template str = {%- if tools %}\n {{- '<|im_start|>...
llama_model_loader: - kv 39: general.quantization_version u32 = 2
llama_model_loader: - kv 40: general.file_type u32 = 24
llama_model_loader: - kv 41: quantize.imatrix.file str = E:\LLM\llama_qwen3_next\imatrix.gguf
llama_model_loader: - kv 42: quantize.imatrix.dataset str = ..\calibration_dataset.txt
llama_model_loader: - kv 43: quantize.imatrix.entries_count u32 = 540
llama_model_loader: - kv 44: quantize.imatrix.chunks_count u32 = 350
llama_model_loader: - type f32: 313 tensors
llama_model_loader: - type q2_K: 20 tensors
llama_model_loader: - type q4_K: 9 tensors
llama_model_loader: - type q6_K: 6 tensors
llama_model_loader: - type iq2_xxs: 10 tensors
llama_model_loader: - type iq3_xxs: 4 tensors
llama_model_loader: - type iq1_s: 14 tensors
llama_model_loader: - type iq4_xs: 8 tensors
llama_model_loader: - type iq1_m: 72 tensors
llama_model_loader: - type bf16: 351 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = IQ1_S - 1.5625 bpw
print_info: file size = 27.20 GiB (2.93 BPW)
load: printing all EOG tokens:
load: - 151643 ('<|endoftext|>')
load: - 151645 ('<|im_end|>')
load: - 151662 ('<|fim_pad|>')
load: - 151663 ('<|repo_name|>')
load: - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch = qwen3next
print_info: vocab_only = 0
print_info: n_ctx_train = 262144
print_info: n_embd = 2048
print_info: n_layer = 48
print_info: n_head = 16
print_info: n_head_kv = 2
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
print_info: n_gqa = 8
print_info: n_embd_k_gqa = 512
print_info: n_embd_v_gqa = 512
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 5120
print_info: n_expert = 512
print_info: n_expert_used = 10
print_info: n_expert_groups = 0
print_info: n_group_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 262144
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 4
print_info: ssm_d_inner = 4096
print_info: ssm_d_state = 128
print_info: ssm_dt_rank = 32
print_info: ssm_n_group = 16
print_info: ssm_dt_b_c_rms = 0
print_info: model type = ?B
print_info: model params = 79.67 B
print_info: general.name = Qwen3 Next 80B A3B Instruct
print_info: vocab type = BPE
print_info: n_vocab = 151936
print_info: n_merges = 151387
print_info: BOS token = 151643 '<|endoftext|>'
print_info: EOS token = 151645 '<|im_end|>'
print_info: EOT token = 151645 '<|im_end|>'
print_info: PAD token = 151643 '<|endoftext|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 151659 '<|fim_prefix|>'
print_info: FIM SUF token = 151661 '<|fim_suffix|>'
print_info: FIM MID token = 151660 '<|fim_middle|>'
print_info: FIM PAD token = 151662 '<|fim_pad|>'
print_info: FIM REP token = 151663 '<|repo_name|>'
print_info: FIM SEP token = 151664 '<|file_sep|>'
print_info: EOG token = 151643 '<|endoftext|>'
print_info: EOG token = 151645 '<|im_end|>'
print_info: EOG token = 151662 '<|fim_pad|>'
print_info: EOG token = 151663 '<|repo_name|>'
print_info: EOG token = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors: CPU_Mapped model buffer size = 25265.84 MiB
load_tensors: Vulkan0 model buffer size = 27246.99 MiB
..........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = false
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host output buffer size = 0.58 MiB
llama_kv_cache: Vulkan0 KV buffer size = 96.00 MiB
llama_kv_cache: size = 96.00 MiB ( 4096 cells, 12 layers, 1/1 seqs), K (f16): 48.00 MiB, V (f16): 48.00 MiB
llama_memory_recurrent: Vulkan0 RS buffer size = 75.38 MiB
llama_memory_recurrent: size = 75.38 MiB ( 1 cells, 48 layers, 1 seqs), R (f32): 3.38 MiB, S (f32): 72.00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context: Vulkan0 compute buffer size = 307.75 MiB
llama_context: Vulkan_Host compute buffer size = 105.01 MiB
llama_context: graph nodes = 9168
llama_context: graph splits = 402 (with bs=512), 438 (with bs=1)
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: added <|repo_name|> logit bias = -inf
common_init_from_params: added <|file_sep|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 8
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | AVX512 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | main: interactive mode on. == Running in interactive mode. ==
🌐 Website Name SuggestionHostVPS.pro (or your brand name) 🧩 1. Website Structure & Pages✅ Homepage (Landing Page)Goal: Convert visitors into leads/customers in <10 seconds. Key Sections:
✅ Features PageGoal: Educate visitors on why VPS beats shared hosting. Sections:
llama_perf_sampler_print: sampling time = 87.45 ms / 708 runs ( 0.12 ms per token, 8096.43 tokens per second) |
|
@lovedheart please stop spamming this topic with irrelevant posts. This is a CPU only implementation focussed on correctness. GPU support will come in followup PRs |
The OP needs to be updated with instructions - we cannot expect people to read all 430 comments in the thread. |
Updated the first post with a README. |
There was a problem hiding this comment.
I tested parallel generations and they seem to work:
make -j && ./bin/llama-parallel -m ../models/qwen3-next/ggml-model-q8_0.gguf -np 5 -ns 8 --temp 0 To move this forward, I suggest doing the following:
- Extract all new ops into a separate PR
- After it is approved, rebase this PR on top of the new ops
- Try to reduce the nodes in the compute graph if possible (I'll take a look into this now)
| ggml_tensor * get_s_l(int32_t il) const; | ||
|
|
||
| int32_t s_copy(int i) const; | ||
| bool has_previous_state() const; |
| #include <mutex> | ||
|
|
||
| // Forward declarations for internal cache access | ||
| struct llama_memory_hybrid; | ||
| struct llama_memory_recurrent; |
| float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); | ||
| float * src = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 ); | ||
| ggml_vec_tri_f32(ne0, i01, dst_ptr, src, keep_org_val, c, ttype); |
There was a problem hiding this comment.
| float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); | |
| float * src = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 ); | |
| ggml_vec_tri_f32(ne0, i01, dst_ptr, src, keep_org_val, c, ttype); | |
| float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1); | |
| float * src_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); | |
| ggml_vec_tri_f32(ne0, i01, dst_ptr, src_ptr, keep_org_val, c, ttype); |
108 participants
EDIT: README FIRST
This is an implementation of a new type of attention gating in GGML.
Therefore, this implementation will be focused on CORRECTNESS ONLY.
Speed tuning and support for more architectures will come in future PRs.
Please do not spam this threads with reports about performance, especially on backend architectures (CUDA, Vulkan).
===
It's been a real learning experience, not gonna lie, but if someone with hybrid model implementation experience (@gabe-l-hart ?) has some quick tips, I'd be grateful.
Currently at the stage of "graph builds, but first decode complains about wrong memory model", probably not building the inputs correctly.
Resolves #15940