llama : initial Mamba-2 support

[compilade]

Follow-up from #8519 (comment). This should fix #7727 and fix #8519.

I've implemented the fully recurrent mode of Mamba-2, because it's very similar to Mamba-1, and also because it seems like the most appropriate mode for text generation.

This does not implement the sequentially semistructured matrix mode, because I'm not yet sure how the block decomposition would fit within the batch and ubatch framework of llama.cpp, and how the chunk size should be chosen. If the recurrent mode is faster at single-user auto-regressive text generation, then I'm not sure how to keep the graph node structure constant when using the most appropriate technique for the batch size.

If the sequentially semistructured matrix mode is eventually implemented, it should help with prompt processing speed for large prompts.

What to expect

(mostly taken from #8519 (comment))

The state in Mamba-2 is bigger than I thought; Mamba-Codestral-7B-v0.1 takes 263.5 MiB (in F32) per sequence (e.g. with -np 1), compared to 38 MiB (also in F32) for Falcon-Mamba-7B (which is based on Mamba-1). But that remains constant whatever the context size. Mamba-2 is easier to implement efficiently, so the bigger state does not really impede inference speed.

However, a big downside right now with recurrent models in llama.cpp is the lack of state rollback (which is implemented through state checkpoints in #7531, but needs to be re-adapted to #8526), so the prompt will be reprocessed a lot if using llama-server. I think using llama-cli in conversation mode does not have this problem, however (or maybe only the bare interactive mode with --in-prefix and --in-suffix, not sure).

This initial implementation is CPU-only, but uses SIMD for the SSM scan, so even though the state is bigger than for Mamba-1 models, in my tests, the speed of Mamba2-130M is similar or better than Mamba-130M (but still not that fast compared to transformer-based models with an empty context), when both are run on CPU.

The speed of Mamba-2 models seems comparable to Transformer-based models when the latter have 2k to 4k tokens in their context.

Summary of changes

• Add support for Mamba2ForCausalLM (including the official Mamba-2 models, and Mamba-Codestral-7B-v0.1)
  • Note that config.json needs to contain "architectures": ["Mamba2ForCausalLM"], for the convert script to properly detect the architecture.
• View Mamba-1 as having d_inner (aka 2 * n_embd) heads of size 1.
  • This simplifies the handling of shapes in ggml_ssm_scan
• ggml
  • Implement Mamba-2's selective state update in ggml_ssm_scan.
    • Re-using the same operator as Mamba-1, because it's pretty much the same operation. (except for how ssm_a is broadcast)
  • Fuse the operation with ssm_d into ggml_ssm_scan
    • Otherwise it would need to be transposed, because the dot-products are done head-wise.
  • Implement Mamba-2's SSM scan with GGML_SIMD.
    • This is possible because there is no element-wise expf in the state update unlike with Mamba-1.
  • Avoid state copies for the SSM state (both for Mamba-1 and Mamba-2) by passing state ids to ggml_ssm_scan.
    • Mamba-2 states are huge. Otherwise masking and copying took close to 10% of the CPU time according to perf.

Other

Here's my favorite quote from Section 3.3 of https://arxiv.org/abs/2405.21060:

Furthermore—by a twist of fate—structured state space models and sequentially semiseparable matrices have the same acronyms, underscoring their equivalence! Conveniently we can use any of these acronyms SSM (state space model or semiseparable matrix), SSS (structured state space or sequentially semiseparable), or SS (state space or semiseparable) interchangeably to unambiguously refer to either concept.

TODO

• Rebase onto master after merging llama : simplify Mamba with advanced batch splits #8526.
• Avoid unnecessary moves of the state
• Adapt the Metal kernels and the tests from ggml : add SSM Metal kernels #8546 to the updated ggml_ssm_scan
• Remove the new GGML_MUL fast broadcast path because it's not used anymore to mask the states.
• Maybe use a new metadata key instead of {arch}.ssm.time_step_rank for the number of heads of Mamba-2, because it's not really the rank of the time step (well, maybe kind of).
  • The meaning of the number of heads and the time-step rank is overlapping enough in Mamba-2 that I think this is fine.
• Maybe not fuse the multiplication with ssm_d in ggml_ssm_scan?
• Maybe split ggml_ssm_scan to separate the implementations for Mamba-1 and Mamba-2, although they do have a lot in common.
  • Seems like they can be distinguished easily enough at the time of kernel dispatch.

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Medium

[ngxson]

Hey @compilade , thanks for implementing this!

I tried converting https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1 using convert_hf_to_gguf.py, but it gives error:

    with open(dir_model / "config.json", "r", encoding="utf-8") as f:
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: 'config.json'

Nevertheless, I successfully converted a Mamba-Codestral transformers-compatible model: https://huggingface.co/Molbap/code2 (Need to comment out the line raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()") in convert_hf_to_gguf.py)

Run it output model (remember to select the correct chat template, since the model does not come with one):

make llama-cli -j && ./llama-cli -m ../models/mcode-7.3B-Q8_0.gguf -cnv -p "You are a helpful assistant" --chat-template mistral -ngl 0

The result looks promising, but I have no idea why there are [UNK_BYTE_0x29681...]. It seems like the there is a problem with space character:

<<SYS>>Youareahelpfulassistant<</SYS>>
> hi
[UNK_BYTE_0xe29681▁Hello]Hello![UNK_BYTE_0xe29681▁How]How[UNK_BYTE_0xe29681▁can]can[UNK_BYTE_0xe29681▁I]I[UNK_BYTE_0xe29681▁assist]assist[UNK_BYTE_0xe29681▁you]you[UNK_BYTE_0xe29681▁today]today?

Link to download GGUF: https://huggingface.co/ngxson/codestral-mamba-llamacpp-test/tree/main

[compilade]

Hey @compilade , thanks for implementing this!

I tried converting https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1 using convert_hf_to_gguf.py, but it gives error:

    with open(dir_model / "config.json", "r", encoding="utf-8") as f:
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: 'config.json'

@ngxson

The steps I took to convert Mamba-Codestral-7B-v0.1 are the following:

1. Rename consolidated.safetensors to model.safetensors
2. Rename params.json to config.json
3. Add the line "architectures": ["Mamba2ForCausalLM"], in config.json
4. Rename tokenizer.model.v3 to tokenizer.model
5. Use convert_hf_to_gguf.py as usual.

I did not have tokenization problems in my tests. Maybe because I was using the original SentencePiece tokenizer instead of a BPE tokenizer.

That tokenizer.json in the transformers-compatible version seems to have problematic spaces. It uses the SentencePiece space escaping instead of the BPE one. Its normalizer seems to revert the escaping, but that's not handled in llama.cpp.

There are probably still problems with the SentencePiece tokenizer too, like the lack of special tokens (control tokens seem to be identified correctly, the only difference seems to be with the 20 [REFERENCE_DOC_{n}] tokens (where n is 0 to 19), which tokenzier.json identifies as non-special added tokens (maps to USER_DEFINED for llama.cpp), while tokenizer.model identifies them as NORMAL tokens).

I think the SentencePiece tokenizer should be preferred for this model; it should be easier to handle without workarounds. I should change that in convert_hf_to_gguf.py. Meanwhile either not include tokenizer.json or rename it to something else.

[ngxson]

Thanks for the guide! I've successfully converted the original repository the gguf by following your steps.

For the transformers-compatible, I will try to contact the one who made it. Hopefully it will be fixed soon.

I'm wondering if convert_hf_to_gguf.py can automatically handle the renaming of params.json, consolidated.safetensors and tokenizer.model.v3? For now, my fear is that someone who use automated tools like gguf-my-repo will be stuck due to this issue.

(Also cc @Vaibhavs10 since he's the maintainer of gguf-my-repo.)

[Vaibhavs10]

Hey @compilade/ @ngxson - JFYI - the transformers weights are now merged in the main repo: https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1

If you face any issues with the conversion with this could you open an issue on the repo for us to track! 🤗

[1ns0mni4c]

Any updates on when Codestral Mamba should be supported?

[learning-chip]

Nice work! Just a note on the ssm_scan kernel performance: a better fused implementation by the flash-linear-attention project can give the equivalent functionality as Mamba2's original kernel: fla-org/flash-linear-attention#49 , and runs 2x faster: fla-org/flash-linear-attention#50

[molbap]

Hi @compilade ! I worked on repo conversion for the transformers-compatible mamba2 version, let us know if you need anything from us to move forward with this PR :)

[HanClinto]

I'm wondering if convert_hf_to_gguf.py can automatically handle the renaming of params.json, consolidated.safetensors and tokenizer.model.v3? For now, my fear is that someone who use automated tools like gguf-my-repo will be stuck due to this issue.

(Also cc @Vaibhavs10 since he's the maintainer of gguf-my-repo.)

It sounds like having a simple fallback of expected filenames would be a reasonable thing to include here? I don't know that we want to maintain a ton of different ones, but adding a second layer of fallbacks for alternate filenames doesn't feel arduous.

[compilade]

It sounds like having a simple fallback of expected filenames would be a reasonable thing to include here? I don't know that we want to maintain a ton of different ones, but adding a second layer of fallbacks for alternate filenames doesn't feel arduous.

@HanClinto

That's not really a problem anymore (at least for Mamba-Codestral) since the official repo was updated in https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1/commit/88085f9cdfa832c3aca8a0315a4520cf7558c947 to use more standard names.

What is currently blocking this is that the Metal and CUDA kernels for ggml_ssm_scan need to be updated BUT before that, I want to refactor the operator to completely avoid copying Mamba-2 states (because otherwise the unnecessary copies use a non-negligible fraction of the memory bandwidth (10% of total text generation inference time on my laptop), since Mamba-2 states are big).

[hg0428]

Any updates on this?

[gabe-l-hart]

Ok, I'm finally back in business with my CUDA box. I've got initial sniff-test results, all looking strong. I'm running off of my GraniteFour branch (which includes all of the changes here).

Model conversion setup

python convert_hf_to_gguf.py /path/to/hf-model
./build/bin/llama-quantize /path/to/hf-model/model-name-F16.gguf Q4_K_M

Model test setup

For each model, I'm testing the grid of CPU/GPU x F16/Q4_K_M

sniff-test.sh

#!/usr/bin/env bash

model_dir=""
prompt="Tell me a story about a developer and their dog"
n_predict="100"

# Parse CLI args
while [[ "$#" -gt 0 ]]; do
  case $1 in
    -m|--model-dir)
      model_dir="$2"
      shift # past argument
      shift # past value
      ;;
    -p|--prompt)
      prompt="$2"
      shift # past argument
      shift # past value
      ;;
    -h|--help)
      echo "Usage: $0 --model-dir <path> [--prompt <string>] [-h | --help]"
      echo ""
      echo "  --model-dir, -m: Parent directory for the model (required)"
      echo "  --prompt, -p: Prompt to test with (optional, default: $prompt)"
      echo "  --n-predict, -n: Number of tokens to predict (optional, default: $n_predict)"
      echo "  -h, --help: Show this help message and exit"
      exit 0
      ;;
    *)
      echo "Unknown parameter passed: $1"
      if [ -z "$model_dir" ]; then
        echo "Error: Missing required argument --model-dir. Use -h for help."
        exit 1
      fi
      ;;
  esac
done

# Check if required arguments were provided
if [ -z "$model_dir" ]; then
  echo "Error: Missing required argument --model-dir. Use -h for help."
  exit 1
fi

echo "Model directory: $model_dir"

if [ -n "$prompt" ]; then
  echo "Prompt: $prompt"
else
  echo "No prompt provided, using an empty string."
fi

f16_gguf=$(find $model_dir -name "*-F16.gguf")
q4_gguf=$(find $model_dir -name "*Q4_K_M.gguf")

echo "F16: $f16_gguf"
echo "Q4_K_M: $a4_gguf"

run_llama_cli() {
  local model=$1
  local use_gpu=$2

  # Set the -ngl flag based on the input bool
  local ngl_value=${use_gpu:='0'}

  # Construct the command with the given arguments
  ./build/bin/llama-cli -m $model --temp 0 -p "$prompt" -n $n_predict -no-cnv -ngl $ngl_value 2>&1 | grep --color=never "llama_perf_"
}

# Run the test grid
echo "------> CPU / F16"
run_llama_cli $f16_gguf 0
echo
echo "------> CPU / Q4_K_M"
run_llama_cli $q4_gguf 0
echo
echo "------> GPU / F16"
run_llama_cli $f16_gguf 999
echo
echo "------> GPU / A4_K_M"
run_llama_cli $q4_gguf 999

(vibe coded with granite3.3 😉)

Sniff test results

./sniff-test.sh -m ~/models/state-spaces/mamba2-2.7b/
Model directory: /home/ghart/models/state-spaces/mamba2-2.7b/
Prompt: Tell me a story about a developer and their dog
F16: /home/ghart/models/state-spaces/mamba2-2.7b/mamba2-2.7B-F16.gguf
Q4_K_M: 
------> CPU / F16
llama_perf_sampler_print:    sampling time =       4.50 ms /   110 runs   (    0.04 ms per token, 24444.44 tokens per second)
llama_perf_context_print:        load time =     424.76 ms
llama_perf_context_print: prompt eval time =     109.02 ms /    10 tokens (   10.90 ms per token,    91.72 tokens per second)
llama_perf_context_print:        eval time =    5892.84 ms /    99 runs   (   59.52 ms per token,    16.80 tokens per second)
llama_perf_context_print:       total time =    6024.16 ms /   109 tokens

------> CPU / Q4_K_M
llama_perf_sampler_print:    sampling time =       5.28 ms /   110 runs   (    0.05 ms per token, 20837.28 tokens per second)
llama_perf_context_print:        load time =     243.90 ms
llama_perf_context_print: prompt eval time =     103.00 ms /    10 tokens (   10.30 ms per token,    97.09 tokens per second)
llama_perf_context_print:        eval time =    3349.81 ms /    99 runs   (   33.84 ms per token,    29.55 tokens per second)
llama_perf_context_print:       total time =    3476.10 ms /   109 tokens

------> GPU / F16
llama_perf_sampler_print:    sampling time =       4.00 ms /   110 runs   (    0.04 ms per token, 27486.26 tokens per second)
llama_perf_context_print:        load time =     975.95 ms
llama_perf_context_print: prompt eval time =      92.11 ms /    10 tokens (    9.21 ms per token,   108.57 tokens per second)
llama_perf_context_print:        eval time =    1324.86 ms /    99 runs   (   13.38 ms per token,    74.73 tokens per second)
llama_perf_context_print:       total time =    1431.39 ms /   109 tokens

------> GPU / A4_K_M
llama_perf_sampler_print:    sampling time =       3.88 ms /   110 runs   (    0.04 ms per token, 28372.45 tokens per second)
llama_perf_context_print:        load time =     351.65 ms
llama_perf_context_print: prompt eval time =      25.90 ms /    10 tokens (    2.59 ms per token,   386.10 tokens per second)
llama_perf_context_print:        eval time =     862.56 ms /    99 runs   (    8.71 ms per token,   114.77 tokens per second)
llama_perf_context_print:       total time =     902.52 ms /   109 tokens

./sniff-test.sh -m ~/models/mistralai/Mamba-Codestral-7B-v0.1/
Model directory: /home/ghart/models/mistralai/Mamba-Codestral-7B-v0.1/
Prompt: Tell me a story about a developer and their dog
F16: /home/ghart/models/mistralai/Mamba-Codestral-7B-v0.1/Mamba-Codestral-7B-v0.1-F16.gguf
Q4_K_M: 
------> CPU / F16
llama_perf_sampler_print:    sampling time =       4.79 ms /   111 runs   (    0.04 ms per token, 23163.61 tokens per second)
llama_perf_context_print:        load time =     823.50 ms
llama_perf_context_print: prompt eval time =     215.69 ms /    11 tokens (   19.61 ms per token,    51.00 tokens per second)
llama_perf_context_print:        eval time =   12671.57 ms /    99 runs   (  128.00 ms per token,     7.81 tokens per second)
llama_perf_context_print:       total time =   12908.29 ms /   110 tokens

------> CPU / Q4_K_M
llama_perf_sampler_print:    sampling time =       5.45 ms /   111 runs   (    0.05 ms per token, 20363.24 tokens per second)
llama_perf_context_print:        load time =     353.57 ms
llama_perf_context_print: prompt eval time =     211.97 ms /    11 tokens (   19.27 ms per token,    51.89 tokens per second)
llama_perf_context_print:        eval time =    6059.11 ms /    99 runs   (   61.20 ms per token,    16.34 tokens per second)
llama_perf_context_print:       total time =    6291.71 ms /   110 tokens

------> GPU / F16
llama_perf_sampler_print:    sampling time =       4.27 ms /   111 runs   (    0.04 ms per token, 26007.50 tokens per second)
llama_perf_context_print:        load time =    2350.44 ms
llama_perf_context_print: prompt eval time =     107.96 ms /    11 tokens (    9.81 ms per token,   101.89 tokens per second)
llama_perf_context_print:        eval time =    2619.14 ms /    99 runs   (   26.46 ms per token,    37.80 tokens per second)
llama_perf_context_print:       total time =    2740.47 ms /   110 tokens

------> GPU / A4_K_M
llama_perf_sampler_print:    sampling time =       4.20 ms /   111 runs   (    0.04 ms per token, 26415.99 tokens per second)
llama_perf_context_print:        load time =     750.08 ms
llama_perf_context_print: prompt eval time =      32.42 ms /    11 tokens (    2.95 ms per token,   339.32 tokens per second)
llama_perf_context_print:        eval time =    1247.74 ms /    99 runs   (   12.60 ms per token,    79.34 tokens per second)
llama_perf_context_print:       total time =    1293.25 ms /   110 tokens

./sniff-test.sh -m ~/models/ibm-granite/granite-4.0-tiny-preview/
Model directory: /home/ghart/models/ibm-granite/granite-4.0-tiny-preview/
Prompt: Tell me a story about a developer and their dog
F16: /home/ghart/models/ibm-granite/granite-4.0-tiny-preview/Granite-4.0-Tiny-Preview-62x915M-F16.gguf
Q4_K_M: 
------> CPU / F16
llama_perf_sampler_print:    sampling time =       4.55 ms /   110 runs   (    0.04 ms per token, 24175.82 tokens per second)
llama_perf_context_print:        load time =     673.76 ms
llama_perf_context_print: prompt eval time =     105.89 ms /    10 tokens (   10.59 ms per token,    94.43 tokens per second)
llama_perf_context_print:        eval time =    3750.80 ms /    99 runs   (   37.89 ms per token,    26.39 tokens per second)
llama_perf_context_print:       total time =    3892.95 ms /   109 tokens

------> CPU / Q4_K_M
llama_perf_sampler_print:    sampling time =       4.34 ms /   110 runs   (    0.04 ms per token, 25345.62 tokens per second)
llama_perf_context_print:        load time =     379.04 ms
llama_perf_context_print: prompt eval time =      70.77 ms /    10 tokens (    7.08 ms per token,   141.30 tokens per second)
llama_perf_context_print:        eval time =    2267.82 ms /    99 runs   (   22.91 ms per token,    43.65 tokens per second)
llama_perf_context_print:       total time =    2374.33 ms /   109 tokens

------> GPU / F16
llama_perf_sampler_print:    sampling time =       3.97 ms /   110 runs   (    0.04 ms per token, 27707.81 tokens per second)
llama_perf_context_print:        load time =    2392.76 ms
llama_perf_context_print: prompt eval time =     151.07 ms /    10 tokens (   15.11 ms per token,    66.19 tokens per second)
llama_perf_context_print:        eval time =    1010.63 ms /    99 runs   (   10.21 ms per token,    97.96 tokens per second)
llama_perf_context_print:       total time =    1190.72 ms /   109 tokens

------> GPU / A4_K_M
llama_perf_sampler_print:    sampling time =       3.97 ms /   110 runs   (    0.04 ms per token, 27686.89 tokens per second)
llama_perf_context_print:        load time =     889.56 ms
llama_perf_context_print: prompt eval time =     210.62 ms /    10 tokens (   21.06 ms per token,    47.48 tokens per second)
llama_perf_context_print:        eval time =     857.46 ms /    99 runs   (    8.66 ms per token,   115.46 tokens per second)
llama_perf_context_print:       total time =    1096.69 ms /   109 tokens

[gabe-l-hart]

I've also verified coherent responses are consistent for each model between CPU/GPU and F16/Q4_K_M (within expected quantization precision limits).

The machine I'm running on has the following setup:

• RHEL 9
• 2x L40S
• CUDA 12.9

[gabe-l-hart]

Running tests with llama-parallel:

$ ./bin/llama-parallel -m /path/to/model.gguf -np 5 -ns 12 --temp 0 --repeat-penalty 1.1 [-pps]

• mamba2-2.7b/ggml-model-Q4_K_M.gguf:

  • w/ pps:
    • Results are coherent and correlated with the prompt
    • It ran the What is the meaning of life? prompt three time. The first two produced identical results, but the third one did not, indicating some nondeterminism related to parallelism and order.
  • w/out pps:
    • Results are coherent and correlated with the prompt
    • The same behavior shows up where different results come up for the same prompt when repeated
    • The results for prompts do differ from running with -pps

• Mamba-Codestral-7B-v0.1/ggml-model-Q4_K_M.gguf:

  • w/ pps:
    • Results are coherent and correlated with the prompt
    • In my run, all responses for repeated prompts were identical (not necessarily conclusive evidence that it can't happen)
  • w/out pps:
    • Results are coherent and correlated with the prompt
    • Repeated prompts do result in different responses

• granite-4.0-tiny-preview/ggml-model-Q4_K_M.gguf:

  • w/ pps
    • Results are coherent and correlated with the prompt
    • Repeated prompts do result in different responses
  • w/out pps
    • SEG FAULT! (debugging time...)

[gabe-l-hart]

I think it's pretty clear that the segfault is something I'm doing wrong with Granite and the hybrid cache, so that shouldn't have any impact on this branch. The only thing from my testing that might be concerning is the inconsistency between results for the same prompt when run using llama-parallel. @compilade would you expect these to always return consistent responses with --temp 0?

[compilade]

The only thing from my testing that might be concerning is the inconsistency between results for the same prompt when run using llama-parallel. @compilade would you expect these to always return consistent responses with --temp 0?

@gabe-l-hart Yes I would expect consistent response; however, when the batch sizes differ (which can happen during continuous batching), the floating-point accumulations in matmuls can be different (because of non-associativity of float addition), which might cause the differences you're seeing (unless the different outputs look broken).

At least in test-model-random (from #14139) I did notice very small (on the order of 1e-14) differences in the outputs for Mamba (which might apply to Mamba-2 too. Right, I tried this in https://github.com/compilade/llama.cpp/tree/compilade/test-model-random-mamba2, and there are small differences on the order of at most 1e-12, which I think are caused by floating-point non-associativity. They are relatively negligible for correctness (for comparison, the same test for a Llama2-like model has differences on the order of 1e-7)).

[gabe-l-hart]

@compilade In digging into the segfault for Granite 4, a member of our team (@AnmolS1) found some thread safety issues in the recurrent cache that were triggering the segfault. He put a std::mutex around apply and prepare and was able to see the issue resolve (draft PR inbound once he completes the checkboxes for IBM OSS contributions). I have two thoughts from what he found:

1. Is it possible that thread safety issues could be contributing to the inconsistent responses above?
2. Are there any thoughts you have on a more nuanced approach to thread safety rather than serializing invocation of apply / prepare?

[compilade]

a member of our team (@AnmolS1) found some thread safety issues in the recurrent cache that were triggering the segfault. He put a std::mutex around apply and prepare and was able to see the issue resolve

@gabe-l-hart This is confusing to me, because I was under the impression that these methods were always called from a single thread (since I think llama_decode should never be called concurrently on the same llama_context? (Although there isn't any mutex preventing that for now, I think the examples/tools still don't call it concurrently, except maybe the server...)). I'm likely wrong here. @ggerganov is llama_decode intended to be callable concurrently on the same llama_context from multiple threads?

Under what conditions did the segfault happen? I assume it's with https://github.com/gabe-l-hart/llama.cpp/tree/GraniteFour? With which tool/example and which args?

I assume it's with ./bin/llama-parallel -m /path/to/model.gguf -np 5 -ns 12 --temp 0 --repeat-penalty 1.1 with https://huggingface.co/ibm-granite/granite-4.0-tiny-preview?

I'm currently downloading granite-4.0-tiny-preview to attempt reproducing the issue.

[ggerganov]

I don't see how there could be a thread safety issue in the llama-parallel example. The apply and prepare should not require a mutex - they are called by the same thread.

[gabe-l-hart]

Ok, interesting. It is indeed happening with this command

./bin/llama-parallel -m ~/models/granite-4.0-tiny-preview/ggml-model-Q4_K_M.gguf -np 5 -ns 12 --temp 0 --repeat-penalty 1.1

It only happens when using the hybrid cache, so it's definitely something about the interplay between the two. When I run with a full debug build, I see this error:

Assertion failed: (status == LLAMA_MEMORY_STATUS_SUCCESS), function apply, file llama-memory-recurrent.cpp, line 1074.
Abort trap: 6

It's happening here when the hybrid cache delegates apply to the child recurrent cache. I tried commenting the assertion out yesterday, but it then fell through to an invalid access trying to use a const llama_ubatch & that was null in the call to find_slot. The interesting thing is that it doesn't seem to happen with -pps, so it seems to have something to do with parallel prefill. I'll dig some more too.

[gabe-l-hart]

I forgot to mention, the status when it hits that assertion is LLAMA_MEMORY_STATUS_NO_UPDATE, so it seems that somehow the recurrent and attention sub-contexts are getting out-of-sync.

[gabe-l-hart]

(sorry if this is mixing PR contexts. We can take this discussion to the Granite PR too if we're pretty confident it isn't causing issues for non-hybrid models)

[gabe-l-hart]

@compilade @ggerganov Now that we've tracked down the source of the hybrid cache seg fault, is there anything holding up this branch?

[ggerganov]

Seem good to merge. Up to @compilade for the final word in case something else is needed before merging.

[compilade]

I've tested most things I wasn't completely sure about (CUDA, SVE), and inference on those platform does seem to work properly for both Mamba-1 and Mamba-2 models, with -ngl 0 and -ngl 99 (and it looks like 0b6f6be also fixes RWKV inference when compiled with SVE on a c7g AWS instance).

Weird small models like https://huggingface.co/delphi-suite/v0-mamba-100k seem to work even when compiled with -DGGML_CUDA=ON since 71bef66 (it failed with an assert previously, but ran correctly in a CPU-only build).

I will merge this without further changes today (Wednesday, July 2nd) around 17:00 UTC (in approx. 8 hours), unless there's an objection or if I find other problems before then.

[gabe-l-hart]

Huge thank you @compilade! It's wonderful to have this fully merged