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 |
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 |
|
@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 |
|
I'm working on imatrix quants, takes a hot minute but I can also make it public before it's fully ready |
ilintar/Qwen3-Next-80B-A3B-Instruct-GGUF - did them a long time ago :) |
|
If someone has interested in low-bit model, may try https://huggingface.co/lovedheart/Qwen3-Next-80B-A3B-Instruct-GGUF. I normally tried at least one real-world problem for testing with a complete llama-cli log attached. Any feedback appreciated. |
|
I'm able to run qwen3next on my 3090+5090, thank you so much! |
|
I just downloaded a source code of llama.cpp and built it, but it seems like i still can not load the model. (Unable to load the model error).; arch linux, unsloth's quantisations |
Just tried with the model I made with vulkan build and it had no error on Fedora with my AMD 395 Can give it a shot to double check: |
|
@M3l-Idk for the record though I tried with unsloth's just now and also got no error, so may be on your end, you got any more info? |
Not really, i dont have time today to debug it. Just wanted to know if its just me or everyone has this problem. Thank you anyway. |
|
@bartowski1182 Im not sure if the right place to ask, but getting "llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'qwen3next'" error with unsloth D:\llama.cpp\llama.cpp>llama-cli -m "D:\llama.cpp\llama.cpp\models\Qwen3-Next-80B-A3B-Instruct-Q2_K.gguf" --n-gpu-layers 30 |
|
@mrfrosty009 You need to rebuild llama.cpp from source! See https://docs.unsloth.ai/models/qwen3-next#llama.cpp-run-qwen3-next-80b-a3b-instruct-tutorial for llama.cpp rebuilding for Unsloth related quants Also the official llama.cpp build docs at https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md is recommended! |
Or, y'know, they could use the actual build instructions from this repo instead of your third-party guide :) |
|
@ddh0 Oh apologies I linked it because they're specifically using the Unsloth quant - but yes using the official llama.cpp build instructions is recommend! I edited my comment as well |
|
@ddh0 @danielhanchen Thanks! Just couple of min ago I noticed that it was linking to old llama folder on another disk, not the one I wanted. After typing right folder, it worked, thanks! |
|
@pwilkin i just wanted to say that you did a great job in this pr |
|
Guys, I genuinely appreciate all the valuable time you've put into this, but for whatever reasons, there is some problem with this implementation. I'm comparing using fastllm, in both instances with an instruct Q4 quant (fastllm uses this one, which doesn't use gguf https://huggingface.co/fastllm/Qwen3-Next-80B-A3B-Instruct-UD-Q4_K_M/tree/main ), same temperature, top_k etc (--temp 0.6 --top-k 20 --top-p 0.95 but there is minium p that I can see in fastllm). When prompted "You must reason step by step, but do NOT output the steps — only the final answer. A dragon has 3 boxes:
The dragon makes three statements:
Based only on these statements, determine with certainty which boxes contain truth or lie. If it cannot be fully determined, output: “Undetermined”. Final answer only." Both implementation expose what they are thinking (they are not the thinking models), but the llama.cpp implementation reaches the conclusion "Undetermined" in 6580 words while the fastllm implementation accomplishes the task in 2235 words (I don't think fastllm exposes the token count). This does not happen on shorter prompts (e.g., joke explanation), so it appears specifically in longer reasoning chains. Not saying it’s “dumb,” but the verbosity difference suggests something may be off internally. |
|
@juanml82 Given that (a) you're doing non-greedy decoding (b) you have not fixed a seed (c) both answers are correct it's much more likely that the difference you're reporting is based on the random seed than on anything in the implementation details. |
|
there is one issue that im having tho, no matter which unsloth's quantization i use, after prompting "create a flappy bird game in html" model gets stuck while trying to code some path to a non existing png file, it just starts hallucinating random numbers as the image's name. |
|
and the code is always the same if i remember correctly |
try https://chat.qwen.ai/, the same issue: 
The model entered an infinite loop and eventually stopped outputting. |
|
CPU-only benchmark: Qwen3-Next vs Qwen3 MoE performance Hardware: AMD EPYC 9454P (48c/96t, Zen 4), DDR5-4800 12-channel, 377GB RAM
Qwen3-Next shows ~5x slower tg than Qwen3-30B-A3B despite both having 3B active parameters. Even gpt-oss-120B with 5B active params runs 3x faster on tg. Thanks for all the effort on this either way! |
|
@jdvpro oh, we'll get to optimizing it on CPU as well, don't worry. |
It seems that pp performance on the CPU should be at least 3 times higher. On the Ryzen AI Max+ 395, I get these results with build: ab6726e (7227) CPU (Vulkan -ngl 0) ./llama-bench -m /home/lynx/llama-swap/models/Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf -fa 1 -r 1 -ngl 0
build: ab6726e (7227) CPU (pure CPU backend) ./llama-bench -m /home/lynx/llama-swap/models/Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf -fa 1
build: ab6726e (7227) Vulkan ./llama-bench -m /home/lynx/llama-swap/models/Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf -fa 1 -r 1 -ngl 999
build: ab6726e (7227) Update: Added more relevant values with pure CPU backend |
|
It's also interesting to note that, thanks to its architecture, Qwen3 Next's performance and tg speeds don't drop as much as Qwen3 30b (Vulkan backend) as the context increases. And starting with a context of 32768, Qwen3 30b doesn't show any significant speed advantages. Vulkan ./llama-bench -m /home/lynx/llama-swap/models/Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf,/home/lynx/llama-swap/models/Qwen3-Coder-30B-A3B-Instruct-UD-Q4_K_XL.gguf -fa 1 -r 1 -ngl 999 -d 512,2048,4096,8192,16384,32768,65536
build: ab6726e (7227) 
However, if Qwen3 Next's performance and tg speeds can be further increased over time through optimizations, that would be incredible. Thank you again to everyone for the opportunity to run Qwen3 Next locally! P.S. Sorry for the spam. Updated: Added graphs for clarity |
|
I dont know if this model is supposed to do that, but it seems like the more text it generates the faster it gets over time, like it can even get twice as fast over time than in the beggining. |
This could happen in the first few runs if the model isn't entirely loaded on GPU as llama.cpp loads the model into RAM. Otherwise it shouldn't happen at all. Text generation at long context get much less slowdown compared to standard attention models, though. |
20 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).
CURRENT STATE: core is done
===
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.
Resolves #15940