improve the vram calculator

[Yuyz0112]

Please check the internal design doc.

[Copilot]

Pull Request Overview

This PR centralizes inference and vLLM parameters into configuration constants, refactors core VRAM calculation functions to use those constants, and removes the CPU resources UI and related CPU calculations.

• Introduce INFERENCE_CONFIG and VLLM_CONFIG to manage empirical parameters
• Refactor calculateKVCache, calculateActivations, calculateFrameworkOverhead, calculateMultiDeviceOverhead, and estimateGenerationSpeed to use new configs
• Remove CPU memory/cores calculation logic and the corresponding UI section

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/pages/vram-calculator/VramCalculatorPage.tsx	Remove CPU resources block; drop unused type & Label imports
src/lib/vram-calculator.ts	Add config constants, refactor calculation functions, remove CPU logic

Comments suppressed due to low confidence (3)

src/lib/vram-calculator.ts:122

• Consider adding unit tests for calculateKVCache to cover scenarios with different effective batch sizes based on VLLM_CONFIG.ACTIVE_USER_RATIO and MEMORY_POOL_OVERHEAD.

function calculateKVCache(

src/lib/vram-calculator.ts:136

• [nitpick] The variable name cachePerToken is somewhat generic; consider renaming it to kvCachePerToken to clarify that it refers to KV cache byte size per token.

  const cachePerToken =

src/pages/vram-calculator/VramCalculatorPage.tsx:557

• Since the CPU resources section has been removed, consider cleaning up related translation keys (e.g., pages.vramCalculator.sections.cpuResources) to avoid unused localization entries.

            <div className="border border-border rounded-lg p-6">

[Copilot]

[nitpick] Consider extracting the GPU memory tier thresholds (80, 40, 24, 16) into a configuration object or enum to improve maintainability and avoid magic numbers.

[samuel032khoury]

OK