Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Sending in tokens one at a time vs all at once gives different logits #4479

Closed
4 tasks done
Phylliida opened this issue Dec 14, 2023 · 6 comments
Closed
4 tasks done

Sending in tokens one at a time vs all at once gives different logits #4479

Phylliida opened this issue Dec 14, 2023 · 6 comments
Labels
bug-unconfirmed stale

Comments

@Phylliida
Copy link

Phylliida commented Dec 14, 2023 •
edited
Loading

Prerequisites

Please answer the following questions for yourself before submitting an issue.

  • I am running the latest code. Development is very rapid so there are no tagged versions as of now.
  • I carefully followed the README.md.
  • I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
  • I reviewed the Discussions, and have a new bug or useful enhancement to share.

Expected Behavior

context [2]
n_past 0

context [2, 3]
n_past 1

Should result in the same logits as

context [2, 3]
n_past 0

(sending all tokens in at the same time)

Current Behavior

They provide different logits. They are similar, but different. I'm assuming this is due to KV-Cache, maybe it can't be fixed because of that. Feel free to close this issue, I just thought it was worth documenting. It may be related to this issue

Steps to Reproduce

Using llama-cpp-python (probably easy enough to do in c++ as well), just run

from llama_cpp import Llama
llm = Llama(model_path="openhermes-2.5-mistral-7b.Q5_K_M.gguf", n_gpu_layers=64, n_ctx=2048)

llm.reset()
llm._ctx.kv_cache_clear()
llm.eval([2])
llm.eval([3])
logitsOneAtATime = llm.eval_logits[0]

llm.reset()
llm._ctx.kv_cache_clear()
llm.eval([2, 3])
logitsAllAtOnce = llm.eval_logits[0]

print("logitsOneAtATime:", logitsOneAtATime[0:5])
print("logitsAllAtOnce:", logitsAllAtOnce[0:5])
print("equal?", np.all(np.array(logitsOneAtATime)==np.array(logitsAllAtOnce)))

Gives

logitsOneAtATime: [0.5897696018218994, 3.2608237266540527, 2.3381338119506836, 9.189224243164062, 0.524137556552887]
logitsAllAtOnce: [1.0115106105804443, 3.683368682861328, 2.2214267253875732, 9.284174919128418, 0.9480904340744019]
equal? False

Failure Log:

I think this behavior has existed for a while, but here's my full outputs:

Full llama-cpp output log 
ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6
llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from Y:/ai/openhermes-2.5-mistral-7b.Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: - tensor    0:                token_embd.weight q5_K     [  4096, 32002,     1,     1 ]
llama_model_loader: - tensor    1:              blk.0.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor    2:              blk.0.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor    3:              blk.0.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor    4:         blk.0.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor    5:            blk.0.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor    6:              blk.0.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor    7:            blk.0.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor    8:           blk.0.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor    9:            blk.0.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   10:              blk.1.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   11:              blk.1.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   12:              blk.1.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   13:         blk.1.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   14:            blk.1.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   15:              blk.1.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   16:            blk.1.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   17:           blk.1.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   18:            blk.1.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   19:              blk.2.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   20:              blk.2.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   21:              blk.2.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   22:         blk.2.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   23:            blk.2.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   24:              blk.2.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   25:            blk.2.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   26:           blk.2.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   27:            blk.2.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   28:              blk.3.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   29:              blk.3.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   30:              blk.3.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   31:         blk.3.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   32:            blk.3.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   33:              blk.3.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   34:            blk.3.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   35:           blk.3.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   36:            blk.3.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   37:              blk.4.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   38:              blk.4.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   39:              blk.4.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   40:         blk.4.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   41:            blk.4.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   42:              blk.4.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   43:            blk.4.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   44:           blk.4.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   45:            blk.4.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   46:              blk.5.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   47:              blk.5.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   48:              blk.5.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   49:         blk.5.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   50:            blk.5.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   51:              blk.5.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   52:            blk.5.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   53:           blk.5.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   54:            blk.5.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   55:              blk.6.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   56:              blk.6.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   57:              blk.6.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   58:         blk.6.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   59:            blk.6.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   60:              blk.6.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   61:            blk.6.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   62:           blk.6.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   63:            blk.6.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   64:              blk.7.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   65:              blk.7.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   66:              blk.7.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   67:         blk.7.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   68:            blk.7.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   69:              blk.7.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   70:            blk.7.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   71:           blk.7.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   72:            blk.7.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   73:              blk.8.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   74:              blk.8.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   75:              blk.8.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   76:         blk.8.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   77:            blk.8.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   78:              blk.8.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   79:            blk.8.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   80:           blk.8.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   81:            blk.8.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   82:              blk.9.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   83:              blk.9.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   84:              blk.9.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   85:         blk.9.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   86:            blk.9.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   87:              blk.9.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   88:            blk.9.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   89:           blk.9.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   90:            blk.9.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   91:             blk.10.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   92:             blk.10.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   93:             blk.10.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor   94:        blk.10.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor   95:           blk.10.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   96:             blk.10.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor   97:           blk.10.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor   98:          blk.10.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor   99:           blk.10.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  100:             blk.11.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  101:             blk.11.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  102:             blk.11.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  103:        blk.11.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  104:           blk.11.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  105:             blk.11.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  106:           blk.11.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  107:          blk.11.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  108:           blk.11.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  109:             blk.12.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  110:             blk.12.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  111:             blk.12.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  112:        blk.12.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  113:           blk.12.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  114:             blk.12.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  115:           blk.12.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  116:          blk.12.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  117:           blk.12.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  118:             blk.13.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  119:             blk.13.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  120:             blk.13.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  121:        blk.13.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  122:           blk.13.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  123:             blk.13.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  124:           blk.13.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  125:          blk.13.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  126:           blk.13.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  127:             blk.14.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  128:             blk.14.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  129:             blk.14.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  130:        blk.14.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  131:           blk.14.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  132:             blk.14.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  133:           blk.14.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  134:          blk.14.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  135:           blk.14.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  136:             blk.15.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  137:             blk.15.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  138:             blk.15.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  139:        blk.15.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  140:           blk.15.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  141:             blk.15.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  142:           blk.15.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  143:          blk.15.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  144:           blk.15.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  145:             blk.16.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  146:             blk.16.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  147:             blk.16.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  148:        blk.16.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  149:           blk.16.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  150:             blk.16.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  151:           blk.16.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  152:          blk.16.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  153:           blk.16.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  154:             blk.17.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  155:             blk.17.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  156:             blk.17.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  157:        blk.17.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  158:           blk.17.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  159:             blk.17.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  160:           blk.17.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  161:          blk.17.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  162:           blk.17.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  163:             blk.18.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  164:             blk.18.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  165:             blk.18.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  166:        blk.18.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  167:           blk.18.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  168:             blk.18.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  169:           blk.18.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  170:          blk.18.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  171:           blk.18.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  172:             blk.19.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  173:             blk.19.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  174:             blk.19.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  175:        blk.19.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  176:           blk.19.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  177:             blk.19.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  178:           blk.19.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  179:          blk.19.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  180:           blk.19.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  181:             blk.20.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  182:             blk.20.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  183:             blk.20.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  184:        blk.20.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  185:           blk.20.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  186:             blk.20.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  187:           blk.20.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  188:          blk.20.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  189:           blk.20.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  190:             blk.21.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  191:             blk.21.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  192:             blk.21.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  193:        blk.21.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  194:           blk.21.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  195:             blk.21.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  196:           blk.21.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  197:          blk.21.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  198:           blk.21.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  199:             blk.22.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  200:             blk.22.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  201:             blk.22.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  202:        blk.22.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  203:           blk.22.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  204:             blk.22.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  205:           blk.22.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  206:          blk.22.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  207:           blk.22.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  208:             blk.23.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  209:             blk.23.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  210:             blk.23.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  211:        blk.23.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  212:           blk.23.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  213:             blk.23.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  214:           blk.23.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  215:          blk.23.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  216:           blk.23.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  217:             blk.24.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  218:             blk.24.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  219:             blk.24.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  220:        blk.24.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  221:           blk.24.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  222:             blk.24.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  223:           blk.24.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  224:          blk.24.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  225:           blk.24.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  226:             blk.25.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  227:             blk.25.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  228:             blk.25.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  229:        blk.25.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  230:           blk.25.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  231:             blk.25.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  232:           blk.25.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  233:          blk.25.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  234:           blk.25.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  235:             blk.26.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  236:             blk.26.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  237:             blk.26.attn_v.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  238:        blk.26.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  239:           blk.26.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  240:             blk.26.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  241:           blk.26.ffn_down.weight q5_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  242:          blk.26.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  243:           blk.26.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  244:             blk.27.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  245:             blk.27.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  246:             blk.27.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  247:        blk.27.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  248:           blk.27.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  249:             blk.27.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  250:           blk.27.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  251:          blk.27.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  252:           blk.27.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  253:             blk.28.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  254:             blk.28.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  255:             blk.28.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  256:        blk.28.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  257:           blk.28.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  258:             blk.28.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  259:           blk.28.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  260:          blk.28.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  261:           blk.28.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  262:             blk.29.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  263:             blk.29.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  264:             blk.29.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  265:        blk.29.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  266:           blk.29.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  267:             blk.29.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  268:           blk.29.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  269:          blk.29.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  270:           blk.29.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  271:             blk.30.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  272:             blk.30.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  273:             blk.30.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  274:        blk.30.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  275:           blk.30.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  276:             blk.30.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  277:           blk.30.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  278:          blk.30.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  279:           blk.30.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  280:             blk.31.attn_q.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  281:             blk.31.attn_k.weight q5_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  282:             blk.31.attn_v.weight q6_K     [  4096,  1024,     1,     1 ]
llama_model_loader: - tensor  283:        blk.31.attn_output.weight q5_K     [  4096,  4096,     1,     1 ]
llama_model_loader: - tensor  284:           blk.31.ffn_gate.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  285:             blk.31.ffn_up.weight q5_K     [  4096, 14336,     1,     1 ]
llama_model_loader: - tensor  286:           blk.31.ffn_down.weight q6_K     [ 14336,  4096,     1,     1 ]
llama_model_loader: - tensor  287:          blk.31.attn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  288:           blk.31.ffn_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  289:               output_norm.weight f32      [  4096,     1,     1,     1 ]
llama_model_loader: - tensor  290:                    output.weight q6_K     [  4096, 32002,     1,     1 ]
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = llama
llama_model_loader: - kv   1:                               general.name str              = teknium_openhermes-2.5-mistral-7b
llama_model_loader: - kv   2:                       llama.context_length u32              = 32768
llama_model_loader: - kv   3:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   4:                          llama.block_count u32              = 32
llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  11:                          general.file_type u32              = 17
llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32002]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32002]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32002]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  16:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  17:                tokenizer.ggml.eos_token_id u32              = 32000
llama_model_loader: - kv  18:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  19:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q5_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
llm_load_vocab: special tokens definition check successful ( 261/32002 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = SPM
llm_load_print_meta: n_vocab          = 32002
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
llm_load_print_meta: n_rot            = 128
llm_load_print_meta: n_gqa            = 4
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: n_ff             = 14336
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 32768
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: model type       = 7B
llm_load_print_meta: model ftype      = mostly Q5_K - Medium
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 4.78 GiB (5.67 BPW)
llm_load_print_meta: general.name     = teknium_openhermes-2.5-mistral-7b
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 32000 '<|im_end|>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: PAD token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.12 MiB
llm_load_tensors: using CUDA for GPU acceleration
llm_load_tensors: mem required  =   86.06 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors: VRAM used: 4807.06 MiB
..................................................................................................
llama_new_context_with_model: n_ctx      = 2048
llama_new_context_with_model: freq_base  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_new_context_with_model: KV self size  =  256.00 MiB, K (f16):  128.00 MiB, V (f16):  128.00 MiB
llama_build_graph: non-view tensors processed: 676/676
llama_new_context_with_model: compute buffer total size = 159.32 MiB
llama_new_context_with_model: VRAM scratch buffer: 156.00 MiB
llama_new_context_with_model: total VRAM used: 4963.07 MiB (model: 4807.06 MiB, context: 156.00 MiB)
AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 0 | VSX = 0 |

Git log:

commit 6744dbe924a317e3e2a5a2a4a2037061b2223449 (HEAD -> master, tag: b1641, origin/master, origin/HEAD)
Author: slaren <slarengh@gmail.com>
Date:   Thu Dec 14 20:05:21 2023 +0100

    ggml : use ggml_row_size where possible (#4472)

    * ggml : use ggml_row_size where possible

    ggml-ci

    * ggml : move ggml_nbytes_split to ggml-cuda.cu

commit cafcd4f89500b8afef722cdb08088eceb8a22572 (tag: b1640)
Author: slaren <slarengh@gmail.com>
Date:   Thu Dec 14 16:52:08 2023 +0100

    ggml : remove n_dims from ggml_tensor (#4469)

    ggml-ci

commit c50e40016394f124b97ce39da48148b1f6c01833
Author: wonjun Jang <strutive07@gmail.com>
Date:   Thu Dec 14 21:44:49 2023 +0900

    py : add protobuf dependency (#4466)

commit 20a68a7030ee06e8eb7eb8e24ae4ac52dc17803f (tag: b1638)
Author: LostRuins <39025047+LostRuins@users.noreply.github.com>
Date:   Thu Dec 14 20:13:33 2023 +0800
@JohannesGaessler
Copy link
Collaborator

Same issue with 0 GPU layers? Does setting the batch size to 1 result in the same logits?

@Phylliida
Copy link
Author

Phylliida commented Dec 21, 2023 •
edited
Loading

The issue only seems to occur when GPU layers>=1 and n_batch>=2

No GPU layers or n_batch=1 makes the issue go away (results in the same logits)

@ggerganov
Copy link
Member

There has been some discussion in the past in this regard: #3014

@JohannesGaessler
Copy link
Collaborator

Generally speaking different logits can be expected between batch size 1 and batch size 2 because they use different matrix multiplications kernels that are going to result in differences in rounding error. I am not observing significant differences in perplexity between batch size 1 and batch size 2. Furthermore if you actually calculate the token probabilities that arise from the logits posted here you find that they are essentially the same:

[1.83458128e-04 2.65192371e-03 1.05400310e-03 9.95938811e-01 1.71804021e-04]
[2.54121506e-04 3.67633126e-03 8.52121086e-04 9.94978921e-01 2.38505494e-04]

The biggest differences are on the first and last token which are not going to be sampled anyways because they get assigned ~0.01% probability, barely more than the ~0.003% probability that each token would have on average. The most likely token goes from 99.5% to 99.4%. My interpretation of what's happening is that those tokens that would make no sense anyways are going to be close to 0 in terms of activations and as such are more susceptible to noise from rounding error. Those tokens therefore can get large relative changes to their probabilities but because they are not going to be sampled from anyways it doesn't matter.

@github-actions GitHub Actions
Copy link
Contributor

This issue is stale because it has been open for 30 days with no activity.

@github-actions github-actions bot added the stale label Mar 18, 2024
@github-actions GitHub Actions
Copy link
Contributor

github-actions bot commented Apr 2, 2024

This issue was closed because it has been inactive for 14 days since being marked as stale.

@github-actions github-actions bot closed this as completed Apr 2, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
bug-unconfirmed stale
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@ggerganov @Phylliida @JohannesGaessler