feat (calculator): Added a hybrid model tab and support V4

[Potterluo]

feat (calculator): Added a hybrid model tab and optimized the KV caching forex interface

• Added two tabs, Standard Model and Hybrid Model, for model selection
• Implemented dedicated configuration options for DeepSeek V4 series hybrid models

[ygwpz]

⚠️ Warning: The MQA formula appears incomplete. It shows head_dim but the calculation logic in performCalculation uses hidden_size * (kvHeads / num_attention_heads) for the else branch. The formula description should clarify that kv_heads = 1 means the head_dim is hidden_size / attn_heads, and the formula should be consistent with the actual calculation.

[Potterluo]

The MQA formula is correct. When kv_heads = 1, head_dim = hidden_size / attn_heads. The formula 2 × layers × tokens × batch × head_dim / tp is mathematically equivalent to 2 × layers × tokens × batch × hidden_size × (1/attn_heads) / tp. The parameter description already clarifies: head_dim: "Head dimension (hidden_size / attn_heads)". No changes needed.

[ygwpz]

⚠️ Warning: The DSA formula is missing batch_size and / tensor_parallelism. The formula shown is KV = layers × tokens × ... but the actual calculation in performCalculation includes batch_size and divides by tp. Please update the formula display to match the actual calculation logic.

[Potterluo]

✅ Fixed. Updated DSA formula to: Single-GPU KV Cache = num_hidden_layers × num_tokens × batch_size × (kv_lora_rank + qk_rope_head_dim + index_head_dim) / tensor_parallelism × dtype_bytes

[ygwpz]

💡 Suggestion: Add validation for the parsed tokens value. If tokens <= 0 or isNaN(tokens), the calculation will produce invalid results. Consider adding input validation similar to calculateKVCache() function:

if (!tokens || isNaN(tokens) || tokens <= 0) {
    displayError('Invalid Input', 'Please enter a valid positive number for tokens.');
    return;
}

[Potterluo]

✅ Fixed. Added input validation for tokens in calculateHybrid() and calculateHybridMaxTokens().

[ygwpz]

💡 Suggestion: Add validation for batch_size and tp similar to the standard model calculation. Invalid values (e.g., tp = 0) would cause division by zero or incorrect results:

if (tp <= 0) {
    displayError('Invalid Input', 'Tensor Parallelism must be at least 1.');
    return;
}

[Potterluo]

✅ Fixed. Added input validation for batch_size and tp to prevent division by zero or invalid results.

[ygwpz]

💡 Suggestion: The hardcoded byte-per-block values (e.g., c4aCompressor: 131072) are magic numbers. Consider adding comments explaining how these values were derived or calculated, to make the code more maintainable and help future readers understand the data source.

[Potterluo]

✅ Fixed. Added calculation source comments explaining how values like c4aCompressor: 131072 are derived (e.g., 512 tokens × 256 KV heads × 1 byte = 131072 B).

[Potterluo]

Thank you for the detailed review! Here are my responses:

---

Comment 1: MQA formula - The MQA formula is correct. When kv_heads = 1, head_dim = hidden_size / attn_heads. The formula 2 × layers × tokens × batch × head_dim / tp is mathematically equivalent to the actual calculation 2 × layers × tokens × batch × hidden_size × (1/attn_heads) / tp. The parameter description already clarifies: head_dim: 'Head dimension (hidden_size / attn_heads)'. No changes needed.

---

Comment 2: DSA formula missing batch_size/tp ✅ Fixed. Updated formula to: Single-GPU KV Cache = num_hidden_layers × num_tokens × batch_size × (kv_lora_rank + qk_rope_head_dim + index_head_dim) / tensor_parallelism × dtype_bytes

---

Comment 3: Input validation for tokens ✅ Fixed. Added validation in calculateHybrid() and calculateHybridMaxTokens().

---

Comment 4: Input validation for batch_size/tp ✅ Fixed. Added complete validation for all input parameters.

---

Comment 5: Magic numbers documentation ✅ Fixed. Added calculation source comments explaining how values like c4aCompressor: 131072 are derived.

---

Please review the latest commit for the fixes.