Skip to content

Fit multimodal chunks once instead of checking every chunk again on evaluation #36

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

Merged
mcharytoniuk merged 1 commit into from
Oct 13, 2026
Merged

Fit multimodal chunks once instead of checking every chunk again on evaluation #36

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
117 changes: 105 additions & 12 deletions llama-cpp-bindings-tests/tests/multimodal_vision.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,8 @@ use llama_cpp_bindings::context::LlamaContext;
use llama_cpp_bindings::error::EvalMultimodalChunksError;
use llama_cpp_bindings::ingest_prompt_chunk::ingest_prompt_chunk;
use llama_cpp_bindings::llama_batch::LlamaBatch;
use llama_cpp_bindings::mtmd::FittedMtmdInputChunk;
use llama_cpp_bindings::mtmd::FittedMtmdInputChunks;
use llama_cpp_bindings::mtmd::MtmdBitmap;
use llama_cpp_bindings::mtmd::MtmdContext;
use llama_cpp_bindings::mtmd::MtmdContextParams;
Expand Down Expand Up @@ -697,6 +699,42 @@ fn eval_chunks_rejects_a_zero_batch_before_evaluating(fixture: &LlamaFixture<'_>
Ok(())
}

#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf"),
n_gpu_layers = 999,
load_mode = Mmap,
n_ctx = 512,
n_batch = 128,
n_ubatch = 64,
mmproj_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "mmproj-F16.gguf"),
)]
#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.6-35B-A3B-GGUF", "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf"),
n_gpu_layers = 999,
load_mode = Mmap,
n_ctx = 512,
n_batch = 128,
n_ubatch = 64,
mmproj_source = HuggingFace("unsloth/Qwen3.6-35B-A3B-GGUF", "mmproj-F16.gguf"),
)]
fn fitting_no_chunks_leaves_nothing_to_evaluate(fixture: &LlamaFixture<'_>) -> Result<()> {
let mtmd_ctx = fixture
.mtmd_context
.expect("mmproj_file declared in attribute");
let llama_ctx = LlamaContext::from_model(
fixture.model,
fixture.backend,
(*fixture.context_params).into_llama_context_params(),
)?;
let fitted_chunks =
FittedMtmdInputChunks::fit(MtmdInputChunks::new()?, mtmd_ctx, llama_ctx.n_ubatch())?;

assert!(fitted_chunks.is_empty());
assert!(FittedMtmdInputChunk::at(&fitted_chunks, 0).is_none());

Ok(())
}

#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf"),
n_gpu_layers = 999,
Expand All @@ -719,13 +757,18 @@ fn eval_single_rejects_a_zero_batch_before_evaluating(fixture: &LlamaFixture<'_>
let mtmd_ctx = fixture
.mtmd_context
.expect("mmproj_file declared in attribute");
let chunks = tokenize_synthetic(fixture, "Describe: <__media__>")?;
let first_chunk = chunks.get(0).context("tokenization produced no chunks")?;
let llama_ctx = LlamaContext::from_model(
fixture.model,
fixture.backend,
(*fixture.context_params).into_llama_context_params(),
)?;
let fitted_chunks = FittedMtmdInputChunks::fit(
tokenize_synthetic(fixture, "Describe: <__media__>")?,
mtmd_ctx,
llama_ctx.n_ubatch(),
)?;
let first_chunk =
FittedMtmdInputChunk::at(&fitted_chunks, 0).context("tokenization produced no chunks")?;

assert_eq!(
first_chunk.eval_single(mtmd_ctx, &llama_ctx, 0, 0, 0, true),
Expand All @@ -736,6 +779,64 @@ fn eval_single_rejects_a_zero_batch_before_evaluating(fixture: &LlamaFixture<'_>
Ok(())
}

#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf"),
n_gpu_layers = 999,
load_mode = Mmap,
n_ctx = 512,
n_batch = 128,
n_ubatch = 64,
mmproj_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "mmproj-F16.gguf"),
)]
#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.6-35B-A3B-GGUF", "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf"),
n_gpu_layers = 999,
load_mode = Mmap,
n_ctx = 512,
n_batch = 128,
n_ubatch = 64,
mmproj_source = HuggingFace("unsloth/Qwen3.6-35B-A3B-GGUF", "mmproj-F16.gguf"),
)]
fn eval_single_rejects_a_micro_batch_smaller_than_the_fitted_one_before_evaluating(
fixture: &LlamaFixture<'_>,
) -> Result<()> {
let mtmd_ctx = fixture
.mtmd_context
.expect("mmproj_file declared in attribute");
let llama_ctx = LlamaContext::from_model(
fixture.model,
fixture.backend,
(*fixture.context_params).into_llama_context_params(),
)?;
let fitted_micro_batch_tokens = llama_ctx.n_ubatch();
let fitted_chunks = FittedMtmdInputChunks::fit(
tokenize_synthetic(fixture, "Describe: <__media__>")?,
mtmd_ctx,
fitted_micro_batch_tokens,
)?;
let first_chunk =
FittedMtmdInputChunk::at(&fitted_chunks, 0).context("tokenization produced no chunks")?;
let smaller_micro_batch_tokens = fitted_micro_batch_tokens - 1;

assert_eq!(
first_chunk.eval_single(
mtmd_ctx,
&llama_ctx,
0,
0,
i32::try_from(smaller_micro_batch_tokens)?,
true
),
Err(MtmdEvalError::EvaluatedMicroBatchSmallerThanFitted {
evaluated_micro_batch_tokens: smaller_micro_batch_tokens,
fitted_micro_batch_tokens,
})
);
assert_eq!(llama_ctx.kv_cache_seq_pos_max(0)?, -1);

Ok(())
}

#[llama_test(
model_source = HuggingFace("unsloth/Qwen3.5-0.8B-GGUF", "Qwen3.5-0.8B-Q4_K_M.gguf"),
n_gpu_layers = 999,
Expand Down Expand Up @@ -1913,7 +2014,7 @@ fn gemma3_eval_chunks_rejects_a_non_causal_image_larger_than_the_micro_batch_bef
n_ubatch = 128,
mmproj_source = HuggingFace("unsloth/gemma-3-4b-it-GGUF", "mmproj-F16.gguf"),
)]
fn gemma3_eval_single_rejects_a_non_causal_image_larger_than_the_micro_batch_before_evaluating(
fn gemma3_fitting_rejects_a_non_causal_image_larger_than_the_micro_batch(
fixture: &LlamaFixture<'_>,
) -> Result<()> {
let mtmd_ctx = fixture
Expand All @@ -1931,22 +2032,14 @@ fn gemma3_eval_single_rejects_a_non_causal_image_larger_than_the_micro_batch_bef
)?;

assert_eq!(
image_chunk.eval_single(
mtmd_ctx,
&llama_ctx,
0,
0,
i32::try_from(llama_ctx.n_batch())?,
true
),
image_chunk.fit_to_micro_batch(mtmd_ctx, llama_ctx.n_ubatch()),
Err(MtmdEvalError::NonCausalChunkExceedsMicroBatch(
NonCausalChunkMicroBatchMismatch {
chunk_tokens: image_chunk.n_tokens(),
micro_batch_tokens: llama_ctx.n_ubatch(),
}
))
);
assert_eq!(llama_ctx.kv_cache_seq_pos_max(0)?, -1);

Ok(())
}
Expand Down
4 changes: 4 additions & 0 deletions llama-cpp-bindings/src/mtmd.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
pub mod fitted_mtmd_input_chunk;
pub mod fitted_mtmd_input_chunks;
pub mod micro_batch_tokens;
pub mod mtmd_bitmap;
pub mod mtmd_bitmap_error;
Expand All @@ -19,6 +21,8 @@ pub mod mtmd_tokenize_error;
pub mod non_causal_chunk_micro_batch_mismatch;
pub mod positive_batch_tokens;

pub use fitted_mtmd_input_chunk::FittedMtmdInputChunk;
pub use fitted_mtmd_input_chunks::FittedMtmdInputChunks;
pub use micro_batch_tokens::micro_batch_tokens;
pub use mtmd_bitmap::MtmdBitmap;
pub use mtmd_bitmap_error::MtmdBitmapError;
Expand Down
Loading
Loading