wehub-resource-sync
76 lines
2.4 KiB
Go
76 lines
2.4 KiB
Go
package vram
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
)
|
|
|
|
// FileInput represents a single model file for estimation (URI and optional pre-known size).
|
|
type FileInput struct {
|
|
URI string
|
|
Size int64
|
|
}
|
|
|
|
// SizeResolver returns the content length in bytes for a given URI.
|
|
type SizeResolver interface {
|
|
ContentLength(ctx context.Context, uri string) (int64, error)
|
|
}
|
|
|
|
// GGUFMeta holds parsed GGUF metadata used for VRAM estimation.
|
|
type GGUFMeta struct {
|
|
BlockCount uint32
|
|
EmbeddingLength uint32
|
|
HeadCount uint32
|
|
HeadCountKV uint32
|
|
MaximumContextLength uint64
|
|
}
|
|
|
|
// GGUFMetadataReader reads GGUF metadata from a URI (e.g. via HTTP Range).
|
|
type GGUFMetadataReader interface {
|
|
ReadMetadata(ctx context.Context, uri string) (*GGUFMeta, error)
|
|
}
|
|
|
|
// EstimateOptions configures VRAM/size estimation.
|
|
// GPULayers and KVQuantBits apply uniformly across all context sizes.
|
|
type EstimateOptions struct {
|
|
GPULayers int
|
|
KVQuantBits int
|
|
}
|
|
|
|
// VRAMAt holds the VRAM estimate at a specific context size.
|
|
type VRAMAt struct {
|
|
ContextLength uint32 `json:"contextLength"`
|
|
VRAMBytes uint64 `json:"vramBytes"`
|
|
VRAMDisplay string `json:"vramDisplay"`
|
|
}
|
|
|
|
// EstimateResult is a flat single-context view of an estimate, suitable for
|
|
// the REST /api/models/vram-estimate response and the MCP vram_estimate tool.
|
|
// It is the legacy shape the LLM and HTTP clients expect (size_bytes /
|
|
// size_display / vram_bytes / vram_display).
|
|
type EstimateResult struct {
|
|
SizeBytes uint64 `json:"size_bytes"`
|
|
SizeDisplay string `json:"size_display"`
|
|
ContextLength uint32 `json:"context_length,omitempty"`
|
|
VRAMBytes uint64 `json:"vram_bytes"`
|
|
VRAMDisplay string `json:"vram_display"`
|
|
}
|
|
|
|
// MultiContextEstimate holds VRAM estimates for one or more context sizes,
|
|
// computed from a single metadata fetch.
|
|
type MultiContextEstimate struct {
|
|
SizeBytes uint64 `json:"sizeBytes"`
|
|
SizeDisplay string `json:"sizeDisplay"`
|
|
Estimates map[string]VRAMAt `json:"estimates"` // keys: context size as string
|
|
ModelMaxContext uint64 `json:"modelMaxContext,omitempty"` // from GGUF metadata
|
|
}
|
|
|
|
// VRAMForContext is a convenience method that returns the VRAMBytes for a
|
|
// specific context size, or 0 if not present.
|
|
func (m MultiContextEstimate) VRAMForContext(ctxLen uint32) uint64 {
|
|
if e, ok := m.Estimates[fmt.Sprint(ctxLen)]; ok {
|
|
return e.VRAMBytes
|
|
}
|
|
return 0
|
|
}
|