Skip to content

Commit 70f1e79

Browse files
mkopcinsmkopcins
committed
cleanup
1 parent 66b3d24 commit 70f1e79

3 files changed

Lines changed: 21 additions & 54 deletions

File tree

‎packages/react-native-executorch/common/runner/encoders/audio_encoder.cpp‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,6 @@
22
#include "audio_encoder.h"
33

44
#include <rnexecutorch/Error.h>
5-
#include <rnexecutorch/Log.h>
65
#include <runner/constants.h>
76

87
#include <executorch/extension/tensor/tensor.h>
@@ -111,10 +110,6 @@ Result<EValue> AudioEncoder::encode(const MultimodalInput &input) {
111110
"audio_encoder output rank=%zd, expected 3",
112111
audio_tensor.dim());
113112
last_token_count_ = static_cast<int32_t>(audio_tensor.size(1));
114-
rnexecutorch::log(rnexecutorch::LOG_LEVEL::Info,
115-
"AudioEncoder: valid_samples=", n_valid,
116-
" padded_samples=", n_padded, " k_blocks=", k_blocks,
117-
" audio_tokens=", last_token_count_);
118113
return exec_result[0];
119114
}
120115

‎packages/react-native-executorch/common/runner/multimodal_decoder_runner.h‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -35,9 +35,6 @@ class MultimodalDecoderRunner : public TextDecoderRunner {
3535
const GenerationConfig &config)
3636
: TextDecoderRunner(module, io_manager, config) {}
3737

38-
// True iff the loaded PTE uses the Gemma-style PLE contract above.
39-
// Reads the kHasPLE constant_method every call; cheap, but callers in
40-
// hot loops should snapshot into a local.
4138
bool has_ple() const {
4239
auto r = module_->get(kHasPLE);
4340
if (r.error() != ::executorch::runtime::Error::Ok) {

‎packages/react-native-executorch/common/runner/multimodal_prefiller.cpp‎

Lines changed: 21 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -8,13 +8,6 @@
88

99
// Ported from executorch/extension/llm/runner/multimodal_prefiller.cpp
1010
// with our token-embedding padding fix and LFM2-VL adaptations.
11-
//
12-
// Supports two PTE shapes, selected from MultimodalDecoderRunner::has_ple()
13-
// (auto-detected at load time):
14-
// * Legacy : token_embedding -> inputs_embeds;
15-
// text_decoder(inputs_embeds, cache_positions).
16-
// * PLE : token_embedding -> (inputs_embeds, ple_tok);
17-
// text_decoder(inputs_embeds, ple_tok, cache_positions).
1811

1912
#include "multimodal_prefiller.h"
2013
#include "constants.h"
@@ -49,17 +42,8 @@ MultimodalPrefiller::prefill(const std::vector<MultimodalInput> &inputs,
4942
"prefill: empty input list");
5043

5144
// ------------------------------------------------------------
52-
// Capacity & shape policy from PTE metadata.
53-
//
54-
// Three knobs drive prefill:
55-
// * get_max_seq_len — text_decoder S cap. In dynamic-shape PTEs this
56-
// is the per-call chunk size (Gemma4 iter201 =
57-
// 128); in static-shape PTEs (LFM2-VL) it is also
58-
// the single-shot prefill cap.
59-
// * get_max_context_len — total KV budget (Gemma4 iter201 = 2048). Only
60-
// materially used by the dynamic-shape path.
61-
// * enable_dynamic_shape — selects between chunked (true) and single-shot
62-
// padded (false) prefill.
45+
// * get_max_seq_len — text_decoder S cap. Max prefill chunk length (<=get_max_conetxt_len)
46+
// * get_max_context_len — total KV budget. Caps max context length for multi-turn conversation.
6347
// ------------------------------------------------------------
6448
int64_t max_seq_len = -1;
6549
{
@@ -92,12 +76,9 @@ MultimodalPrefiller::prefill(const std::vector<MultimodalInput> &inputs,
9276
// ------------------------------------------------------------
9377
// Pass 1: build a fused input_ids buffer spanning all inputs.
9478
//
95-
// Mirrors gemma_export/experiments/infer_image.py::prefill_single_shot:
96-
// llm_ids = prefix_ids + [0] * num_soft + suffix_ids
97-
// Image positions use pad_token_id=0, matching HF modeling_gemma4.py:2190
98-
// (placeholder_id is rewritten to 0 before PLE lookup). The decoder embeds
99-
// at those positions are then overwritten with the vision encoder output
100-
// in pass 2.
79+
// Image positions use pad_token_id (placeholder_id is rewritten to 0 before PLE lookup).
80+
// The decoder embeds at those positions are then
81+
// overwritten with the vision encoder output in pass 2.
10182
// ------------------------------------------------------------
10283
struct ImageSlot {
10384
const MultimodalInput *input; // non-owning, valid for duration of call
@@ -112,12 +93,7 @@ MultimodalPrefiller::prefill(const std::vector<MultimodalInput> &inputs,
11293
// just a TensorImpl*; `Method::get_output(i)` returns `const EValue&` to
11394
// Method-internal storage and Module::execute copies that EValue into the
11495
// returned vector. The copy shares the underlying TensorImpl, so a later
115-
// execute() on the same method — a second audio input in this prefill,
116-
// a Vulkan backend output-buffer reuse across methods, or a load-time
117-
// warm-up — mutates `sizes()` in place under our feet. The original error
118-
// ("audio encoder returned 96 tokens, expected 60") is exactly this:
119-
// slot.num_audio was captured from the FIRST encode, slot.encoded.size(1)
120-
// reflected the SECOND. Mirrors main_mm.cpp:604-675's copy-on-encode.
96+
// execute() on the same method — a second audio input in this prefill would overwrite.
12197
struct AudioSlot {
12298
std::vector<uint8_t> bytes;
12399
::executorch::aten::ScalarType dtype;
@@ -171,23 +147,22 @@ MultimodalPrefiller::prefill(const std::vector<MultimodalInput> &inputs,
171147
AudioSlot{std::move(bytes), audio_tensor.scalar_type(),
172148
static_cast<int64_t>(ids.size()), num_audio, audio_hidden});
173149
ids.insert(ids.end(), static_cast<size_t>(num_audio), 0);
174-
} else if (input.is_text() || input.is_tokens()) {
175-
std::vector<uint64_t> tokens;
176-
if (input.is_text()) {
177-
auto encode_result = tokenizer_->encode(input.get_text());
178-
if (!encode_result.ok()) {
179-
ET_LOG(Error, "Tokenizer encode error %d",
180-
static_cast<uint32_t>(encode_result.error()));
181-
return Error::InvalidArgument;
182-
}
183-
tokens = std::move(*encode_result);
184-
} else {
185-
tokens = input.get_tokens();
150+
} else if (input.is_text()) {
151+
auto encode_result = tokenizer_->encode(input.get_text());
152+
if (!encode_result.ok()) {
153+
ET_LOG(Error, "Tokenizer encode error %d",
154+
static_cast<uint32_t>(encode_result.error()));
155+
return Error::InvalidArgument;
186156
}
157+
std::vector<uint64_t> tokens = std::move(*encode_result);
158+
for (auto t : tokens) {
159+
ids.push_back(static_cast<int64_t>(t));
160+
}
161+
} else if (input.is_tokens()) {
162+
std::vector<uint64_t> tokens = input.get_tokens();
187163
for (auto t : tokens) {
188164
ids.push_back(static_cast<int64_t>(t));
189165
}
190-
} else {
191166
ET_LOG(Error, "Unsupported MultimodalInput type");
192167
return Error::NotSupported;
193168
}
@@ -383,12 +358,12 @@ MultimodalPrefiller::prefill(const std::vector<MultimodalInput> &inputs,
383358
// ------------------------------------------------------------
384359
// Chunked text_decoder calls.
385360
//
386-
// Some PTEs (Gemma4 iter201) hard-cap text_decoder's S dim at
361+
// Some PTEs (Gemma4) hard-cap text_decoder's S dim at
387362
// get_max_seq_len (128) while the prefill budget extends to
388363
// get_max_context_len (2048). KV cache state persists across calls via the
389364
// absolute input_pos vector, so chunking is functionally transparent to
390-
// the model. For single-shot static-shape PTEs (LFM2-VL) chunk_cap ==
391-
// total_len so the loop iterates exactly once — preserving prior behavior.
365+
// the model. For single-shot PTEs chunk_cap == total_len
366+
// so the loop iterates exactly once — preserving prior behavior.
392367
// ------------------------------------------------------------
393368
const int64_t chunk_cap =
394369
decoder_chunk_size > 0 ? decoder_chunk_size : total_len;

0 commit comments

Comments
 (0)