[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M74Z19YFAQE0CJG0XZTP34C6":3,"skill-same:01M74Z19YFAQE0CJG0XZTP34C6":72},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":30,"commitSha":31,"skillDirectory":32,"forksCount":33,"repositoryHtmlUrl":34,"repositoryHomepageUrl":35,"ownerHtmlUrl":36,"ownerBlogUrl":37,"ownerDescription":38,"updatedAt":39,"install":40,"resources":53},"01M74Z19YFAQE0CJG0XZTP34C6","moe-training",null,"Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. Use when training large-scale models with limited compute (5× cost reduction vs dense models), implementing sparse architectures like Mixtral 8x7B or DeepSeek-V3, or scaling model capacity without proportional compute increase. Covers MoE architectures, routing mechanisms, load balancing, expert parallelism, and inference optimization.","KalarisLabs\u002Fresearch-agent-skills","KalarisLabs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F300878568?v=4","01M74Y7MT23GXXC56VM6TVSXZY","Kalarislabs","01M74Y7MTQM9REXZX4SNXWX95C",132,"MIT","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: moe-training\ndescription: Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. Use when training large-scale models with limited compute (5× cost reduction vs dense models), implementing sparse architectures like Mixtral 8x7B or DeepSeek-V3, or scaling model capacity without proportional compute increase. Covers MoE architectures, routing mechanisms, load balancing, expert parallelism, and inference optimization.\nlicense: MIT\nmetadata:\n  version: 1.0.0\n  category: multimodal-and-emerging\n  maintainer: Kalaris Labs\n  tags: Emerging Techniques, MoE, Mixture Of Experts, Sparse Models, DeepSpeed, Expert Parallelism, Mixtral, DeepSeek, Routing, Load Balancing, Efficient Training\n  dependencies: deepspeed, transformers, torch, accelerate\n---\n\n# MoE Training: Mixture of Experts\n\n## When to Use This Skill\n\nUse MoE Training when you need to:\n- **Train larger models** with limited compute (5× cost reduction vs dense models)\n- **Scale model capacity** without proportional compute increase\n- **Achieve better performance** per compute budget than dense models\n- **Specialize experts** for different domains\u002Ftasks\u002Flanguages\n- **Reduce inference latency** with sparse activation (only 13B\u002F47B params active in Mixtral)\n- **Implement SOTA models** like Mixtral 8x7B, DeepSeek-V3, Switch Transformers\n\n**Notable MoE Models**: Mixtral 8x7B (Mistral AI), DeepSeek-V3, Switch Transformers (Google), GLaM (Google), NLLB-MoE (Meta)\n\n## Installation\n\n```bash\n# DeepSpeed with MoE support\npip install deepspeed>=0.6.0\n\n# Megatron-DeepSpeed for large-scale training\ngit clone https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FMegatron-DeepSpeed\ncd Megatron-DeepSpeed\npip install -r requirements.txt\n\n# Alternative: HuggingFace Transformers\npip install transformers accelerate\n```\n\n## Quick Start\n\n### Basic MoE Architecture\n\n```python\nimport torch\nimport torch.nn as nn\n\nclass MoELayer(nn.Module):\n    \"\"\"Sparse Mixture of Experts layer.\"\"\"\n\n    def __init__(self, hidden_size, num_experts=8, top_k=2):\n        super().__init__()\n        self.num_experts = num_experts\n        self.top_k = top_k\n\n        # Expert networks (FFN)\n        self.experts = nn.ModuleList([\n            nn.Sequential(\n                nn.Linear(hidden_size, 4 * hidden_size),\n                nn.GELU(),\n                nn.Linear(4 * hidden_size, hidden_size)\n            )\n            for _ in range(num_experts)\n        ])\n\n        # Gating network (router)\n        self.gate = nn.Linear(hidden_size, num_experts)\n\n    def forward(self, x):\n        # x shape: (batch_size, seq_len, hidden_size)\n        batch_size, seq_len, hidden_size = x.shape\n\n        # Flatten for routing\n        x_flat = x.view(-1, hidden_size)  # (batch_size * seq_len, hidden_size)\n\n        # Compute gate scores\n        gate_logits = self.gate(x_flat)  # (batch_size * seq_len, num_experts)\n\n        # Top-k routing\n        gate_scores = torch.softmax(gate_logits, dim=-1)\n        topk_scores, topk_indices = torch.topk(gate_scores, self.top_k, dim=-1)\n\n        # Normalize top-k scores\n        topk_scores = topk_scores \u002F topk_scores.sum(dim=-1, keepdim=True)\n\n        # Dispatch and combine expert outputs\n        output = torch.zeros_like(x_flat)\n\n        for i in range(self.top_k):\n            expert_idx = topk_indices[:, i]\n            expert_scores = topk_scores[:, i].unsqueeze(-1)\n\n            # Route tokens to experts\n            for expert_id in range(self.num_experts):\n                mask = (expert_idx == expert_id)\n                if mask.any():\n                    expert_input = x_flat[mask]\n                    expert_output = self.experts[expert_id](expert_input)\n                    output[mask] += expert_scores[mask] * expert_output\n\n        # Reshape back\n        return output.view(batch_size, seq_len, hidden_size)\n```\n\n### DeepSpeed MoE Training\n\n```bash\n# Training script with MoE\ndeepspeed pretrain_gpt_moe.py \\\n  --num-layers 24 \\\n  --hidden-size 1024 \\\n  --num-attention-heads 16 \\\n  --seq-length 2048 \\\n  --max-position-embeddings 2048 \\\n  --micro-batch-size 4 \\\n  --global-batch-size 256 \\\n  --train-iters 500000 \\\n  --lr 0.0001 \\\n  --min-lr 0.00001 \\\n  --lr-decay-style cosine \\\n  --num-experts 128 \\\n  --moe-expert-parallel-size 4 \\\n  --moe-loss-coeff 0.01 \\\n  --moe-train-capacity-factor 1.25 \\\n  --moe-eval-capacity-factor 2.0 \\\n  --fp16 \\\n  --deepspeed_config ds_config.json\n```\n\n## Core Concepts\n\n### 1. MoE Architecture\n\n**Key Components:**\n- **Experts**: Multiple specialized FFN networks (typically 8-128)\n- **Router\u002FGate**: Learned network that selects which experts to use\n- **Top-k Routing**: Activate only k experts per token (k=1 or k=2)\n- **Load Balancing**: Ensure even expert utilization\n\n```\nInput Token\n    ↓\nRouter (Gate Network)\n    ↓\nTop-k Expert Selection (e.g., 2 out of 8)\n    ↓\nExpert 1 (weight: 0.6) + Expert 5 (weight: 0.4)\n    ↓\nWeighted Combination\n    ↓\nOutput\n```\n\n### 2. Routing Mechanisms\n\n**Top-1 Routing (Switch Transformer):**\n```python\n# Simplest routing: one expert per token\ngate_logits = router(x)  # (batch, seq_len, num_experts)\nexpert_idx = torch.argmax(gate_logits, dim=-1)  # Hard routing\n```\n\n**Top-2 Routing (Mixtral):**\n```python\n# Top-2: two experts per token\ngate_scores = torch.softmax(router(x), dim=-1)\ntop2_scores, top2_indices = torch.topk(gate_scores, k=2, dim=-1)\n\n# Normalize scores\ntop2_scores = top2_scores \u002F top2_scores.sum(dim=-1, keepdim=True)\n\n# Combine expert outputs\noutput = (top2_scores[:, :, 0:1] * expert_outputs[top2_indices[:, :, 0]] +\n          top2_scores[:, :, 1:2] * expert_outputs[top2_indices[:, :, 1]])\n```\n\n**Expert Choice Routing:**\n```python\n# Experts choose top-k tokens (instead of tokens choosing experts)\n# Guarantees perfect load balancing\nexpert_scores = router(x).transpose(-1, -2)  # (batch, num_experts, seq_len)\ntopk_tokens = torch.topk(expert_scores, k=capacity_per_expert, dim=-1)\n```\n\n### 3. Load Balancing\n\n**Auxiliary Loss:**\n```python\ndef load_balancing_loss(gate_logits, expert_indices, num_experts):\n    \"\"\"Encourage uniform expert usage.\"\"\"\n    # Fraction of tokens routed to each expert\n    expert_counts = torch.bincount(expert_indices.flatten(), minlength=num_experts)\n    expert_fraction = expert_counts.float() \u002F expert_indices.numel()\n\n    # Gate probability for each expert (average across tokens)\n    gate_probs = torch.softmax(gate_logits, dim=-1).mean(dim=0)\n\n    # Auxiliary loss: encourage alignment\n    aux_loss = num_experts * (expert_fraction * gate_probs).sum()\n\n    return aux_loss\n\n# Add to main loss\ntotal_loss = language_model_loss + 0.01 * load_balancing_loss(...)\n```\n\n**Router Z-Loss (Stability):**\n```python\ndef router_z_loss(logits):\n    \"\"\"Encourage router to have lower entropy (more decisive).\"\"\"\n    z_loss = torch.logsumexp(logits, dim=-1).pow(2).mean()\n    return z_loss\n\ntotal_loss = lm_loss + 0.01 * aux_loss + 0.001 * router_z_loss(gate_logits)\n```\n\n### 4. Expert Parallelism\n\n```python\n# DeepSpeed configuration\n{\n  \"train_batch_size\": 256,\n  \"fp16\": {\"enabled\": true},\n  \"moe\": {\n    \"enabled\": true,\n    \"num_experts\": 128,\n    \"expert_parallel_size\": 8,  # Distribute 128 experts across 8 GPUs\n    \"capacity_factor\": 1.25,    # Expert capacity = tokens_per_batch * capacity_factor \u002F num_experts\n    \"drop_tokens\": true,        # Drop tokens exceeding capacity\n    \"use_residual\": false\n  }\n}\n```\n\n## Training Configuration\n\n### DeepSpeed MoE Config\n\n```json\n{\n  \"train_batch_size\": 256,\n  \"gradient_accumulation_steps\": 1,\n  \"optimizer\": {\n    \"type\": \"Adam\",\n    \"params\": {\n      \"lr\": 0.0001,\n      \"betas\": [0.9, 0.999],\n      \"eps\": 1e-8\n    }\n  },\n  \"fp16\": {\n    \"enabled\": true,\n    \"loss_scale\": 0,\n    \"initial_scale_power\": 16\n  },\n  \"moe\": {\n    \"enabled\": true,\n    \"num_experts\": 128,\n    \"expert_parallel_size\": 8,\n    \"moe_loss_coeff\": 0.01,\n    \"train_capacity_factor\": 1.25,\n    \"eval_capacity_factor\": 2.0,\n    \"min_capacity\": 4,\n    \"drop_tokens\": true,\n    \"use_residual\": false,\n    \"use_tutel\": false\n  },\n  \"zero_optimization\": {\n    \"stage\": 1\n  }\n}\n```\n\n### Training Script\n\n```bash\n#!\u002Fbin\u002Fbash\n\n# Mixtral-style MoE training\ndeepspeed --num_gpus 8 pretrain_moe.py \\\n  --model-parallel-size 1 \\\n  --num-layers 32 \\\n  --hidden-size 4096 \\\n  --num-attention-heads 32 \\\n  --seq-length 2048 \\\n  --max-position-embeddings 4096 \\\n  --micro-batch-size 2 \\\n  --global-batch-size 256 \\\n  --train-iters 500000 \\\n  --save-interval 5000 \\\n  --eval-interval 1000 \\\n  --eval-iters 100 \\\n  --lr 0.0001 \\\n  --min-lr 0.00001 \\\n  --lr-decay-style cosine \\\n  --lr-warmup-iters 2000 \\\n  --clip-grad 1.0 \\\n  --weight-decay 0.1 \\\n  --num-experts 8 \\\n  --moe-expert-parallel-size 4 \\\n  --moe-loss-coeff 0.01 \\\n  --moe-train-capacity-factor 1.25 \\\n  --moe-eval-capacity-factor 2.0 \\\n  --disable-moe-token-dropping \\\n  --fp16 \\\n  --deepspeed \\\n  --deepspeed_config ds_config_moe.json \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n## Advanced Patterns\n\nDetails, code examples and parameter tables: [references\u002Fadvanced-patterns.md](references\u002Fadvanced-patterns.md). Read it when this step applies.\n\n## Best Practices\n\n### 1. Expert Count Selection\n\n```python\n# Rule of thumb: More experts = more capacity, but diminishing returns\n# Typical configurations:\n# - Small models (1B-7B): 8-16 experts\n# - Medium models (7B-30B): 8-64 experts\n# - Large models (30B+): 64-256 experts\n\n# Example: Mixtral 8x7B\n# Total params: 47B (8 experts × 7B each)\n# Active params: 13B (2 experts × 7B, top-2 routing)\n# Efficiency: 47B capacity with 13B compute\n```\n\n### 2. Capacity Factor Tuning\n\n```python\n# Capacity = (tokens_per_batch \u002F num_experts) * capacity_factor\n\n# Training: Lower capacity (faster, drops some tokens)\ntrain_capacity_factor = 1.25  # 25% buffer\n\n# Evaluation: Higher capacity (no dropping)\neval_capacity_factor = 2.0    # 100% buffer\n\n# Formula:\nexpert_capacity = int((seq_len * batch_size \u002F num_experts) * capacity_factor)\n```\n\n### 3. Learning Rate Guidelines\n\n```python\n# MoE models need lower LR than dense models\n# - Dense model: lr = 6e-4\n# - MoE model: lr = 1e-4 (3-6× lower)\n\n# Also extend decay schedule\ndense_lr_decay_iters = 300000\nmoe_lr_decay_iters = 500000  # 1.5-2× longer\n```\n\n### 4. Loss Coefficient Tuning\n\n```python\n# Start with standard values\nmoe_loss_coeff = 0.01    # Auxiliary loss (load balancing)\nrouter_z_loss_coeff = 0.001  # Router entropy (stability)\n\n# If load imbalance persists, increase aux loss\nif max_expert_usage \u002F min_expert_usage > 2.0:\n    moe_loss_coeff = 0.1  # Stronger load balancing\n\n# If training unstable, increase z-loss\nif grad_norm > 10.0:\n    router_z_loss_coeff = 0.01\n```\n\n### 5. Avoid Common Pitfalls\n\n```python\n# ❌ Bad: Using same LR as dense model\noptimizer = Adam(model.parameters(), lr=6e-4)\n\n# ✅ Good: Lower LR for MoE\noptimizer = Adam([\n    {'params': model.non_moe_params, 'lr': 6e-4},\n    {'params': model.moe_params, 'lr': 1e-4}\n])\n\n# ❌ Bad: No load balancing\nloss = lm_loss\n\n# ✅ Good: Add auxiliary loss\nloss = lm_loss + 0.01 * aux_loss + 0.001 * z_loss\n\n# ❌ Bad: Too many experts for small dataset\nnum_experts = 128  # Overfitting risk\n\n# ✅ Good: Match experts to data diversity\nnum_experts = 8  # Better for small datasets\n```\n\n## Inference Optimization\n\n### Sparse Inference\n\n```python\n# Only activate top-k experts (huge memory savings)\n@torch.no_grad()\ndef moe_inference(x, model, top_k=2):\n    \"\"\"Sparse MoE inference: only load k experts.\"\"\"\n    # Router\n    gate_logits = model.gate(x)\n    topk_scores, topk_indices = torch.topk(\n        torch.softmax(gate_logits, dim=-1),\n        k=top_k,\n        dim=-1\n    )\n\n    # Load and run only top-k experts\n    output = torch.zeros_like(x)\n    for i in range(top_k):\n        expert_idx = topk_indices[:, i]\n        # Load expert from disk\u002Foffload if needed\n        expert = model.load_expert(expert_idx)\n        output += topk_scores[:, i:i+1] * expert(x)\n\n    return output\n```\n\n## Resources\n\n- **DeepSpeed MoE Tutorial**: https:\u002F\u002Fwww.deepspeed.ai\u002Ftutorials\u002Fmixture-of-experts-nlg\u002F\n- **Mixtral Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.04088\n- **Switch Transformers**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.03961\n- **HuggingFace MoE Guide**: https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fmoe\n- **NVIDIA MoE Blog**: https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fapplying-mixture-of-experts-in-llm-architectures\u002F\n\n## See Also\n\n- `references\u002Farchitectures.md` - MoE model architectures (Mixtral, Switch, DeepSeek-V3)\n- `references\u002Ftraining.md` - Advanced training techniques and optimization\n- `references\u002Finference.md` - Production deployment and serving patterns\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm GPU memory, model checkpoint and licence, framework versions and input formats.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Run inference on a single sample at low resolution or short length first.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Inspect outputs qualitatively and with task metrics; check licence restrictions for generated content.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| Out-of-memory on large inputs | Reduce resolution or sequence length, use half precision, or tile the input. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Check model and dataset licences before commercial or published use.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n","# MoE Training: Mixture of Experts\n\n## When to Use This Skill\n\nUse MoE Training when you need to:\n- **Train larger models** with limited compute (5× cost reduction vs dense models)\n- **Scale model capacity** without proportional compute increase\n- **Achieve better performance** per compute budget than dense models\n- **Specialize experts** for different domains\u002Ftasks\u002Flanguages\n- **Reduce inference latency** with sparse activation (only 13B\u002F47B params active in Mixtral)\n- **Implement SOTA models** like Mixtral 8x7B, DeepSeek-V3, Switch Transformers\n\n**Notable MoE Models**: Mixtral 8x7B (Mistral AI), DeepSeek-V3, Switch Transformers (Google), GLaM (Google), NLLB-MoE (Meta)\n\n## Installation\n\n```bash\n# DeepSpeed with MoE support\npip install deepspeed>=0.6.0\n\n# Megatron-DeepSpeed for large-scale training\ngit clone https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002FMegatron-DeepSpeed\ncd Megatron-DeepSpeed\npip install -r requirements.txt\n\n# Alternative: HuggingFace Transformers\npip install transformers accelerate\n```\n\n## Quick Start\n\n### Basic MoE Architecture\n\n```python\nimport torch\nimport torch.nn as nn\n\nclass MoELayer(nn.Module):\n    \"\"\"Sparse Mixture of Experts layer.\"\"\"\n\n    def __init__(self, hidden_size, num_experts=8, top_k=2):\n        super().__init__()\n        self.num_experts = num_experts\n        self.top_k = top_k\n\n        # Expert networks (FFN)\n        self.experts = nn.ModuleList([\n            nn.Sequential(\n                nn.Linear(hidden_size, 4 * hidden_size),\n                nn.GELU(),\n                nn.Linear(4 * hidden_size, hidden_size)\n            )\n            for _ in range(num_experts)\n        ])\n\n        # Gating network (router)\n        self.gate = nn.Linear(hidden_size, num_experts)\n\n    def forward(self, x):\n        # x shape: (batch_size, seq_len, hidden_size)\n        batch_size, seq_len, hidden_size = x.shape\n\n        # Flatten for routing\n        x_flat = x.view(-1, hidden_size)  # (batch_size * seq_len, hidden_size)\n\n        # Compute gate scores\n        gate_logits = self.gate(x_flat)  # (batch_size * seq_len, num_experts)\n\n        # Top-k routing\n        gate_scores = torch.softmax(gate_logits, dim=-1)\n        topk_scores, topk_indices = torch.topk(gate_scores, self.top_k, dim=-1)\n\n        # Normalize top-k scores\n        topk_scores = topk_scores \u002F topk_scores.sum(dim=-1, keepdim=True)\n\n        # Dispatch and combine expert outputs\n        output = torch.zeros_like(x_flat)\n\n        for i in range(self.top_k):\n            expert_idx = topk_indices[:, i]\n            expert_scores = topk_scores[:, i].unsqueeze(-1)\n\n            # Route tokens to experts\n            for expert_id in range(self.num_experts):\n                mask = (expert_idx == expert_id)\n                if mask.any():\n                    expert_input = x_flat[mask]\n                    expert_output = self.experts[expert_id](expert_input)\n                    output[mask] += expert_scores[mask] * expert_output\n\n        # Reshape back\n        return output.view(batch_size, seq_len, hidden_size)\n```\n\n### DeepSpeed MoE Training\n\n```bash\n# Training script with MoE\ndeepspeed pretrain_gpt_moe.py \\\n  --num-layers 24 \\\n  --hidden-size 1024 \\\n  --num-attention-heads 16 \\\n  --seq-length 2048 \\\n  --max-position-embeddings 2048 \\\n  --micro-batch-size 4 \\\n  --global-batch-size 256 \\\n  --train-iters 500000 \\\n  --lr 0.0001 \\\n  --min-lr 0.00001 \\\n  --lr-decay-style cosine \\\n  --num-experts 128 \\\n  --moe-expert-parallel-size 4 \\\n  --moe-loss-coeff 0.01 \\\n  --moe-train-capacity-factor 1.25 \\\n  --moe-eval-capacity-factor 2.0 \\\n  --fp16 \\\n  --deepspeed_config ds_config.json\n```\n\n## Core Concepts\n\n### 1. MoE Architecture\n\n**Key Components:**\n- **Experts**: Multiple specialized FFN networks (typically 8-128)\n- **Router\u002FGate**: Learned network that selects which experts to use\n- **Top-k Routing**: Activate only k experts per token (k=1 or k=2)\n- **Load Balancing**: Ensure even expert utilization\n\n```\nInput Token\n    ↓\nRouter (Gate Network)\n    ↓\nTop-k Expert Selection (e.g., 2 out of 8)\n    ↓\nExpert 1 (weight: 0.6) + Expert 5 (weight: 0.4)\n    ↓\nWeighted Combination\n    ↓\nOutput\n```\n\n### 2. Routing Mechanisms\n\n**Top-1 Routing (Switch Transformer):**\n```python\n# Simplest routing: one expert per token\ngate_logits = router(x)  # (batch, seq_len, num_experts)\nexpert_idx = torch.argmax(gate_logits, dim=-1)  # Hard routing\n```\n\n**Top-2 Routing (Mixtral):**\n```python\n# Top-2: two experts per token\ngate_scores = torch.softmax(router(x), dim=-1)\ntop2_scores, top2_indices = torch.topk(gate_scores, k=2, dim=-1)\n\n# Normalize scores\ntop2_scores = top2_scores \u002F top2_scores.sum(dim=-1, keepdim=True)\n\n# Combine expert outputs\noutput = (top2_scores[:, :, 0:1] * expert_outputs[top2_indices[:, :, 0]] +\n          top2_scores[:, :, 1:2] * expert_outputs[top2_indices[:, :, 1]])\n```\n\n**Expert Choice Routing:**\n```python\n# Experts choose top-k tokens (instead of tokens choosing experts)\n# Guarantees perfect load balancing\nexpert_scores = router(x).transpose(-1, -2)  # (batch, num_experts, seq_len)\ntopk_tokens = torch.topk(expert_scores, k=capacity_per_expert, dim=-1)\n```\n\n### 3. Load Balancing\n\n**Auxiliary Loss:**\n```python\ndef load_balancing_loss(gate_logits, expert_indices, num_experts):\n    \"\"\"Encourage uniform expert usage.\"\"\"\n    # Fraction of tokens routed to each expert\n    expert_counts = torch.bincount(expert_indices.flatten(), minlength=num_experts)\n    expert_fraction = expert_counts.float() \u002F expert_indices.numel()\n\n    # Gate probability for each expert (average across tokens)\n    gate_probs = torch.softmax(gate_logits, dim=-1).mean(dim=0)\n\n    # Auxiliary loss: encourage alignment\n    aux_loss = num_experts * (expert_fraction * gate_probs).sum()\n\n    return aux_loss\n\n# Add to main loss\ntotal_loss = language_model_loss + 0.01 * load_balancing_loss(...)\n```\n\n**Router Z-Loss (Stability):**\n```python\ndef router_z_loss(logits):\n    \"\"\"Encourage router to have lower entropy (more decisive).\"\"\"\n    z_loss = torch.logsumexp(logits, dim=-1).pow(2).mean()\n    return z_loss\n\ntotal_loss = lm_loss + 0.01 * aux_loss + 0.001 * router_z_loss(gate_logits)\n```\n\n### 4. Expert Parallelism\n\n```python\n# DeepSpeed configuration\n{\n  \"train_batch_size\": 256,\n  \"fp16\": {\"enabled\": true},\n  \"moe\": {\n    \"enabled\": true,\n    \"num_experts\": 128,\n    \"expert_parallel_size\": 8,  # Distribute 128 experts across 8 GPUs\n    \"capacity_factor\": 1.25,    # Expert capacity = tokens_per_batch * capacity_factor \u002F num_experts\n    \"drop_tokens\": true,        # Drop tokens exceeding capacity\n    \"use_residual\": false\n  }\n}\n```\n\n## Training Configuration\n\n### DeepSpeed MoE Config\n\n```json\n{\n  \"train_batch_size\": 256,\n  \"gradient_accumulation_steps\": 1,\n  \"optimizer\": {\n    \"type\": \"Adam\",\n    \"params\": {\n      \"lr\": 0.0001,\n      \"betas\": [0.9, 0.999],\n      \"eps\": 1e-8\n    }\n  },\n  \"fp16\": {\n    \"enabled\": true,\n    \"loss_scale\": 0,\n    \"initial_scale_power\": 16\n  },\n  \"moe\": {\n    \"enabled\": true,\n    \"num_experts\": 128,\n    \"expert_parallel_size\": 8,\n    \"moe_loss_coeff\": 0.01,\n    \"train_capacity_factor\": 1.25,\n    \"eval_capacity_factor\": 2.0,\n    \"min_capacity\": 4,\n    \"drop_tokens\": true,\n    \"use_residual\": false,\n    \"use_tutel\": false\n  },\n  \"zero_optimization\": {\n    \"stage\": 1\n  }\n}\n```\n\n### Training Script\n\n```bash\n#!\u002Fbin\u002Fbash\n\n# Mixtral-style MoE training\ndeepspeed --num_gpus 8 pretrain_moe.py \\\n  --model-parallel-size 1 \\\n  --num-layers 32 \\\n  --hidden-size 4096 \\\n  --num-attention-heads 32 \\\n  --seq-length 2048 \\\n  --max-position-embeddings 4096 \\\n  --micro-batch-size 2 \\\n  --global-batch-size 256 \\\n  --train-iters 500000 \\\n  --save-interval 5000 \\\n  --eval-interval 1000 \\\n  --eval-iters 100 \\\n  --lr 0.0001 \\\n  --min-lr 0.00001 \\\n  --lr-decay-style cosine \\\n  --lr-warmup-iters 2000 \\\n  --clip-grad 1.0 \\\n  --weight-decay 0.1 \\\n  --num-experts 8 \\\n  --moe-expert-parallel-size 4 \\\n  --moe-loss-coeff 0.01 \\\n  --moe-train-capacity-factor 1.25 \\\n  --moe-eval-capacity-factor 2.0 \\\n  --disable-moe-token-dropping \\\n  --fp16 \\\n  --deepspeed \\\n  --deepspeed_config ds_config_moe.json \\\n  --data-path \u002Fpath\u002Fto\u002Fdata \\\n  --vocab-file \u002Fpath\u002Fto\u002Fvocab.json \\\n  --merge-file \u002Fpath\u002Fto\u002Fmerges.txt\n```\n\n## Advanced Patterns\n\nDetails, code examples and parameter tables: [references\u002Fadvanced-patterns.md](references\u002Fadvanced-patterns.md). Read it when this step applies.\n\n## Best Practices\n\n### 1. Expert Count Selection\n\n```python\n# Rule of thumb: More experts = more capacity, but diminishing returns\n# Typical configurations:\n# - Small models (1B-7B): 8-16 experts\n# - Medium models (7B-30B): 8-64 experts\n# - Large models (30B+): 64-256 experts\n\n# Example: Mixtral 8x7B\n# Total params: 47B (8 experts × 7B each)\n# Active params: 13B (2 experts × 7B, top-2 routing)\n# Efficiency: 47B capacity with 13B compute\n```\n\n### 2. Capacity Factor Tuning\n\n```python\n# Capacity = (tokens_per_batch \u002F num_experts) * capacity_factor\n\n# Training: Lower capacity (faster, drops some tokens)\ntrain_capacity_factor = 1.25  # 25% buffer\n\n# Evaluation: Higher capacity (no dropping)\neval_capacity_factor = 2.0    # 100% buffer\n\n# Formula:\nexpert_capacity = int((seq_len * batch_size \u002F num_experts) * capacity_factor)\n```\n\n### 3. Learning Rate Guidelines\n\n```python\n# MoE models need lower LR than dense models\n# - Dense model: lr = 6e-4\n# - MoE model: lr = 1e-4 (3-6× lower)\n\n# Also extend decay schedule\ndense_lr_decay_iters = 300000\nmoe_lr_decay_iters = 500000  # 1.5-2× longer\n```\n\n### 4. Loss Coefficient Tuning\n\n```python\n# Start with standard values\nmoe_loss_coeff = 0.01    # Auxiliary loss (load balancing)\nrouter_z_loss_coeff = 0.001  # Router entropy (stability)\n\n# If load imbalance persists, increase aux loss\nif max_expert_usage \u002F min_expert_usage > 2.0:\n    moe_loss_coeff = 0.1  # Stronger load balancing\n\n# If training unstable, increase z-loss\nif grad_norm > 10.0:\n    router_z_loss_coeff = 0.01\n```\n\n### 5. Avoid Common Pitfalls\n\n```python\n# ❌ Bad: Using same LR as dense model\noptimizer = Adam(model.parameters(), lr=6e-4)\n\n# ✅ Good: Lower LR for MoE\noptimizer = Adam([\n    {'params': model.non_moe_params, 'lr': 6e-4},\n    {'params': model.moe_params, 'lr': 1e-4}\n])\n\n# ❌ Bad: No load balancing\nloss = lm_loss\n\n# ✅ Good: Add auxiliary loss\nloss = lm_loss + 0.01 * aux_loss + 0.001 * z_loss\n\n# ❌ Bad: Too many experts for small dataset\nnum_experts = 128  # Overfitting risk\n\n# ✅ Good: Match experts to data diversity\nnum_experts = 8  # Better for small datasets\n```\n\n## Inference Optimization\n\n### Sparse Inference\n\n```python\n# Only activate top-k experts (huge memory savings)\n@torch.no_grad()\ndef moe_inference(x, model, top_k=2):\n    \"\"\"Sparse MoE inference: only load k experts.\"\"\"\n    # Router\n    gate_logits = model.gate(x)\n    topk_scores, topk_indices = torch.topk(\n        torch.softmax(gate_logits, dim=-1),\n        k=top_k,\n        dim=-1\n    )\n\n    # Load and run only top-k experts\n    output = torch.zeros_like(x)\n    for i in range(top_k):\n        expert_idx = topk_indices[:, i]\n        # Load expert from disk\u002Foffload if needed\n        expert = model.load_expert(expert_idx)\n        output += topk_scores[:, i:i+1] * expert(x)\n\n    return output\n```\n\n## Resources\n\n- **DeepSpeed MoE Tutorial**: https:\u002F\u002Fwww.deepspeed.ai\u002Ftutorials\u002Fmixture-of-experts-nlg\u002F\n- **Mixtral Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.04088\n- **Switch Transformers**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2101.03961\n- **HuggingFace MoE Guide**: https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fmoe\n- **NVIDIA MoE Blog**: https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fapplying-mixture-of-experts-in-llm-architectures\u002F\n\n## See Also\n\n- `references\u002Farchitectures.md` - MoE model architectures (Mixtral, Switch, DeepSeek-V3)\n- `references\u002Ftraining.md` - Advanced training techniques and optimization\n- `references\u002Finference.md` - Production deployment and serving patterns\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm GPU memory, model checkpoint and licence, framework versions and input formats.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Run inference on a single sample at low resolution or short length first.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Inspect outputs qualitatively and with task metrics; check licence restrictions for generated content.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| Out-of-memory on large inputs | Reduce resolution or sequence length, use half precision, or tile the input. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Check model and dataset licences before commercial or published use.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n",{"name":5,"license":15,"metadata":24,"description":7},{"tags":25,"version":26,"category":27,"maintainer":28,"dependencies":29},"Emerging Techniques, MoE, Mixture Of Experts, Sparse Models, DeepSpeed, Expert Parallelism, Mixtral, DeepSeek, Routing, Load Balancing, Efficient Training","1.0.0","multimodal-and-emerging","Kalaris Labs","deepspeed, transformers, torch, accelerate","main","29fe36d26127205057d5e0ed8425f6b0e9a8c13a","skills\u002Fmoe-training",16,"https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills","https:\u002F\u002Fdocs.kalarislabs.com\u002Fresearch-agent-skills","https:\u002F\u002Fgithub.com\u002FKalarisLabs","kalarislabs.com","Applied AI lab building the infrastructure for the agentic era","2026-11-14 17:49:08",{"source":8,"skillName":5,"commands":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add KalarisLabs\u002Fresearch-agent-skills --skill moe-training",{"tool":46,"command":47},"bunx","bunx skills add KalarisLabs\u002Fresearch-agent-skills --skill moe-training",{"tool":49,"command":50},"pnpm","pnpm dlx skills add KalarisLabs\u002Fresearch-agent-skills --skill moe-training","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002FSKILL.md",true,[54,60,64,68],{"path":55,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Fadvanced-patterns.md","REFERENCE","参考文档",2630,"https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002Freferences\u002Fadvanced-patterns.md",{"path":61,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":62,"sourceHtmlUrl":63},"references\u002Farchitectures.md",14044,"https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002Freferences\u002Farchitectures.md",{"path":65,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":66,"sourceHtmlUrl":67},"references\u002Finference.md",8239,"https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002Freferences\u002Finference.md",{"path":69,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":70,"sourceHtmlUrl":71},"references\u002Ftraining.md",9707,"https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fmoe-training\u002Freferences\u002Ftraining.md",[73,80,87,94,101,108],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M74Z8MGT305R11SNY5G28QH5","plos","Prepare manuscripts for PLOS journals (PLOS ONE, PLOS Biology, PLOS Computational Biology, PLOS Genetics, PLOS Medicine, PLOS Pathogens, PLOS Neglected Tropical Diseases, PLOS Global Public Health and others), covering structure, abstract and author summary, Vancouver-style references, the mandatory data availability policy, ethics and financial disclosure statements, figure requirements (PACE), the PLOS LaTeX template, reporting guidelines and PLOS ONE's publication criteria. Use when targeting any PLOS journal or checking compliance with PLOS policies.","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fplos\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M74Z8CCN9APXYHNQ95V1EBGT","pkpd-modeling","Pharmacokinetic and pharmacodynamic modelling and simulation - non-compartmental analysis, compartmental and population PK, PK\u002FPD and exposure-response, TMDD, PBPK orientation, bioequivalence, allometric scaling and first-in-human dose, drug interaction prediction, and Bayesian therapeutic drug monitoring. Use when analysing concentration-time data, deriving exposure metrics, fitting PK or PD models, or evaluating dosing regimens. Triggers include \"pharmacokinetics\", \"pharmacodynamics\", \"PK\u002FPD\", \"NCA\", \"non-compartmental\", \"AUC\", \"Cmax\", \"lambda z\", \"half-life\", \"clearance\", \"volume of distribution\", \"compartmental model\", \"population PK\", \"popPK\", \"NONMEM\", \"nlmixr2\", \"Pharmpy\", \"Monolix\", \"exposure-response\", \"Emax\", \"EC50\", \"indirect response\", \"effect compartment\", \"TMDD\", \"PBPK\", \"bioequivalence\", \"RSABE\", \"ABEL\", \"allometric scaling\", \"first-in-human\", \"MABEL\", \"drug-drug interaction\", \"DDI\", \"ICH M12\", \"concentration-QTc\", \"therapeutic drug monitoring\", \"MIPD\", and \"dosing regimen\".","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fpkpd-modeling\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M74Z86YHJJ1FM3X87BHN76KZ","pinecone","Guides use of Pinecone, a managed serverless vector database, through its Python client and the LangChain and LlamaIndex integrations. Covers creating indexes, upserting and querying vectors, metadata filtering, namespaces, hybrid dense and sparse search, index management, and deleting vectors. Use when building a production RAG system on a hosted vector store, adding semantic search or recommendations without running infrastructure, isolating per-user or per-tenant data with namespaces, combining dense and sparse vectors in one query, or filtering results by metadata. Do not use for self-hosted or local stores (Chroma, Weaviate) or offline similarity search (FAISS).","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fpinecone\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M74Z80T6E89SAJ8MYDD0767Q","pi-agent","Build with and use Pi, the minimal terminal coding harness. Use for installing Pi, configuring providers\u002Fmodels\u002Fsettings\u002Fenvironment variables, creating Pi skills\u002Fextensions\u002Fpackages\u002Fthemes\u002Fprompt templates, embedding Pi through the SDK, integrating over RPC or JSON event streams, parsing sessions, running local models through the llama.cpp router, developing custom Pi providers and TUI components, or using ecosystem packages such as pi-subagents (delegation\u002Forchestration), pi-mcp-adapter (MCP servers), pi-interview (interactive forms), and pi-web-access (web search, fetching, video understanding).","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fpi-agent\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":19},"01M74Z7WF1Y8Q3VFAQPPPRAFMV","phoenix-observability","Open-source AI observability platform for LLM tracing, evaluation, and monitoring. Use when debugging LLM applications with detailed traces, running evaluations on datasets, or monitoring production AI systems with real-time insights.","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fphoenix-observability\u002FSKILL.md",[],[],{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":114,"publishedAt":19},"01M74Z7HNEH1A2AVYR3G1X2QV6","peft-fine-tuning","Fine-tunes LLMs with Hugging Face PEFT, using LoRA, QLoRA, IA3, AdaLoRA, prefix tuning, and prompt tuning so that under 1% of parameters are trained. Covers rank, alpha, and target module selection, loading and merging adapters, multi-adapter serving, and integration with TRL SFTTrainer, Axolotl, and vLLM. Use when fine-tuning 7B-70B models on limited GPU memory, when running QLoRA on a single 24GB GPU, when choosing LoRA rank and alpha, when merging or swapping adapters on one base model, or when debugging CUDA OOM or an adapter that does not apply. Not for full fine-tuning of small models under 1B parameters or cases needing all weights updated.","https:\u002F\u002Fgithub.com\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fblob\u002Fmain\u002Fskills\u002Fpeft-fine-tuning\u002FSKILL.md",[],[]]