Skip to content

fix(GgufInsights): correct KV cache VRAM estimate for quantized types #608

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

Open
andreinknv wants to merge 1 commit into
base: master
Choose a base branch
from
+13 −2
Open

fix(GgufInsights): correct KV cache VRAM estimate for quantized types #608

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
15 changes: 13 additions & 2 deletions src/gguf/insights/GgufInsights.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -603,8 +603,19 @@ export class GgufInsights {
const nEmbdHeadK = this._ggufFileInfo.architectureMetadata.attention?.key_length ?? ((nHead == 0) ? 0 : (nEmbd / nHead));
const nHeadKv: number | number[] = this._ggufFileInfo.architectureMetadata.attention?.head_count_kv ?? nHead;
const nEmbdHeadV = this._ggufFileInfo.architectureMetadata.attention?.value_length ?? ((nHead == 0) ? 0 : nEmbd / nHead);
// `getTypeSizeForGgmlType` returns bytes per BLOCK (matches
// `ggml_type_size` in llama.cpp). For block-quantized types
// (Q4_0, Q8_0, ...) one block holds N elements (block size > 1),
// so per-element bytes = block-bytes / block-elements. F16 / F32
// are scalar (blockSize=1) so the division is a no-op there.
// Without this division, quantized KV-cache estimates overshoot
// by ~32× and the configuration resolver rejects valid configs.
const keyTypeSize = this._llama._bindings.getTypeSizeForGgmlType(kvCacheKeyType) ?? this._llama._consts.ggmlTypeF16Size;
const valueTypeSize = this._llama._bindings.getTypeSizeForGgmlType(kvCacheValueType) ?? this._llama._consts.ggmlTypeF16Size;
const keyBlockSize = this._llama._bindings.getBlockSizeForGgmlType(kvCacheKeyType) ?? 1;
const valueBlockSize = this._llama._bindings.getBlockSizeForGgmlType(kvCacheValueType) ?? 1;
const keyBytesPerElement = keyTypeSize / Math.max(1, keyBlockSize);
const valueBytesPerElement = valueTypeSize / Math.max(1, valueBlockSize);

// source: `llama_model::load_tensors` in `llama-model.cpp`
// repeating layers are assigned to GPU from `i_gpu_start = n_layer + 1 - n_gpu_layers`
Expand Down Expand Up @@ -644,9 +655,9 @@ export class GgufInsights {
}

const gpuKVCacheSize = usingGpu
? ((gpuKvElementsK * keyTypeSize) + (gpuKvElementsV * valueTypeSize))
? ((gpuKvElementsK * keyBytesPerElement) + (gpuKvElementsV * valueBytesPerElement))
: 0;
const cpuKVCacheSize = (cpuKvElementsK * keyTypeSize) + (cpuKvElementsV * valueTypeSize);
const cpuKVCacheSize = (cpuKvElementsK * keyBytesPerElement) + (cpuKvElementsV * valueBytesPerElement);

const recurrentCellSize = Math.max(1, sequences);
const gpuRecurrentStateSize = usingGpu
Expand Down