Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Get Started Free →

MoE Training: Mixture of Experts · openlair | DecimalAI

AI/MLby openlair

MoE Training: Mixture of Experts

Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace.

Never made a case worse

SkillScore

SkillScore91/100

Test results 91

Safety

3/3 safety checks passed

Code scan
Intent
Content

What it changes

73%91%of test cases passed +18 pts

Without skill16 of 22

With skill20 of 22

4 fixed16 passed either way2 still failing0 broke

Measured on gemini-3.6-flash8/7/2026v121 comparable

Track record

Agent runs: 0; All time

See one task, with and without the skill

One test taskDuring MoE training, monitoring shows that max_expert_usage / min_expert_usage exceeds 2.5, indicating severe routing imbalance despite having an auxiliary loss coeff of 0.01. What adjustment to the auxiliary loss coefficient should be made?

Without the skillMissed

With the skillPassed

Same model, same task. Answers are capped and redacted.

Use this skill

Your agent runs on

Adds every file the skill has, pinned so you can update it. No account needed.

pip install "decimalai>=0.15"
decimalai skills add openlair-moe-training
claude --plugin-dir "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents"

The same install from Python, for a notebook or a setup script.

Install

pip install "decimalai>=0.15"

Add the skill

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

print("Added. Start Claude Code with the command below.")

Start Claude Code with it

claude --plugin-dir "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.agents"

Type these in a Claude Code session. No account or API key. It writes the skill and any files it bundles, and asks first when its safety rating isn't Passed.

Add the DecimalAI plugin, once

/plugin marketplace add decimal-labs/decimalai-claude-plugins
/plugin install decimalai@decimalai

Install this skill

/decimalai:install openlair-moe-training

The plugin needs python3 and curl on your machine.

No Python or account. Claude Code picks it up in its next session.

Save it into your project

curl -fsSL --create-dirs -o .claude/skills/openlair-moe-training/SKILL.md https://app.decimal.ai/s/openlair-moe-training/SKILL.md

Copy SKILL.md

Save it in your project, as

.claude/skills/openlair-moe-training/SKILL.md

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Claude Code docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[claude-agent-sdk]"
decimalai skills add openlair-moe-training

Add it to your agent

import asyncio

from claude_agent_sdk import ClaudeAgentOptions
from decimalai.claude_agent_sdk import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

async def main():
    question = "Hello!"
    async with support.run(question) as run:  # one user message, one run
        async for message in run.query(options=ClaudeAgentOptions(max_turns=6)):
            print(message)

# In a notebook or async code: await main()
asyncio.run(main())

Prefer its own loader? ClaudeAgentOptions(plugins=[skills_plugin()], skills="all") reads the same folder. Use one or the other: a skill delivered both ways reaches the prompt twice.

The same install from Python, for a notebook or a setup script. agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[claude-agent-sdk]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import asyncio

from claude_agent_sdk import ClaudeAgentOptions
from decimalai.claude_agent_sdk import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

async def main():
    question = "Hello!"
    async with support.run(question) as run:  # one user message, one run
        async for message in run.query(options=ClaudeAgentOptions(max_turns=6)):
            print(message)

# In a notebook or async code: await main()
asyncio.run(main())

Prefer its own loader? ClaudeAgentOptions(plugins=[skills_plugin()], skills="all") reads the same folder. Use one or the other: a skill delivered both ways reaches the prompt twice.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

import asyncio
from pathlib import Path
from claude_agent_sdk import ClaudeAgentOptions, query

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
options = ClaudeAgentOptions(system_prompt=f"You are a helpful agent.\n\n{skill}", max_turns=6)

async def main():
    async for message in query(prompt="Hello!", options=options):
        print(message)

# In a notebook or async code: await main()
asyncio.run(main())

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Claude Agent SDK docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[openai-agents]"
decimalai skills add openlair-moe-training

Add it to your agent

from agents import Agent, Runner
from decimalai.openai_agents import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = Agent(
    name=support.name,
    instructions=support.instructions,  # your prompt, then the skill list
    tools=[*support.tools],             # your tools, then load_skill
)
# In a notebook or async code: result = await Runner.run(agent, "Hello!")
result = Runner.run_sync(agent, "Hello!")

Using SandboxAgent? capabilities=[Shell(), decimalai.openai_agents.sandbox_skills()] points its own loader at the same folder.

Install

pip install "decimalai[openai-agents]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from agents import Agent, Runner
from decimalai.openai_agents import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = Agent(
    name=support.name,
    instructions=support.instructions,  # your prompt, then the skill list
    tools=[*support.tools],             # your tools, then load_skill
)
# In a notebook or async code: result = await Runner.run(agent, "Hello!")
result = Runner.run_sync(agent, "Hello!")

Using SandboxAgent? capabilities=[Shell(), decimalai.openai_agents.sandbox_skills()] points its own loader at the same folder.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from agents import Agent, Runner

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = Agent(name="support", instructions=f"You are a helpful agent.\n\n{skill}")
# In a notebook or async code: result = await Runner.run(agent, "Hello!")
result = Runner.run_sync(agent, "Hello!")

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

OpenAI Agents docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[openai]"
decimalai skills add openlair-moe-training

Add it to your agent

import json

from openai import OpenAI
from decimalai.openai import agent_config

client = OpenAI()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")


def lookup_order(order_id: str) -> str:
    return f"Order {order_id} shipped yesterday."


# Your tools, as Chat Completions takes them, and what answers each.
TOOLS = [{"type": "function", "function": {
    "name": "lookup_order",
    "description": "Look up the status of an order.",
    "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}},
                   "required": ["order_id"]},
}}]
FUNCTIONS = {"lookup_order": lookup_order}

question = "Where is order A1001?"
with support.run(question) as run:  # one user message, one run
    # run.system is your prompt, then the skill list.
    messages = [{"role": "system", "content": run.system}, run.message]
    while True:
        reply = client.chat.completions.create(
            model="gpt-5.4-mini",
            messages=messages,
            tools=[*TOOLS, *run.tools],  # your tools, then load_skill
        ).choices[0].message
        messages.append(reply)
        if not reply.tool_calls:
            break
        for call in reply.tool_calls:
            # load_skill is answered by the run; your own tools by FUNCTIONS.
            function = FUNCTIONS.get(call.function.name)
            messages.append(run.tool_result(call) or {
                "role": "tool", "tool_call_id": call.id,
                "content": function(**json.loads(call.function.arguments)) if function
                else f"There is no tool named {call.function.name!r}.",
            })
print(reply.content)

# The Responses API takes run.response_tools instead of run.tools.

Install

pip install "decimalai[openai]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import json

from openai import OpenAI
from decimalai.openai import agent_config

client = OpenAI()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")


def lookup_order(order_id: str) -> str:
    return f"Order {order_id} shipped yesterday."


# Your tools, as Chat Completions takes them, and what answers each.
TOOLS = [{"type": "function", "function": {
    "name": "lookup_order",
    "description": "Look up the status of an order.",
    "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}},
                   "required": ["order_id"]},
}}]
FUNCTIONS = {"lookup_order": lookup_order}

question = "Where is order A1001?"
with support.run(question) as run:  # one user message, one run
    # run.system is your prompt, then the skill list.
    messages = [{"role": "system", "content": run.system}, run.message]
    while True:
        reply = client.chat.completions.create(
            model="gpt-5.4-mini",
            messages=messages,
            tools=[*TOOLS, *run.tools],  # your tools, then load_skill
        ).choices[0].message
        messages.append(reply)
        if not reply.tool_calls:
            break
        for call in reply.tool_calls:
            # load_skill is answered by the run; your own tools by FUNCTIONS.
            function = FUNCTIONS.get(call.function.name)
            messages.append(run.tool_result(call) or {
                "role": "tool", "tool_call_id": call.id,
                "content": function(**json.loads(call.function.arguments)) if function
                else f"There is no tool named {call.function.name!r}.",
            })
print(reply.content)

# The Responses API takes run.response_tools instead of run.tools.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from openai import OpenAI

client = OpenAI()
skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
reply = client.chat.completions.create(
    model="gpt-5.4-mini",
    messages=[
        {"role": "system", "content": f"You are a helpful agent.\n\n{skill}"},
        {"role": "user", "content": "Hello!"},
    ],
).choices[0].message

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

OpenAI SDK docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[anthropic]"
decimalai skills add openlair-moe-training

Add it to your agent

import anthropic
from decimalai.anthropic import agent_config

client = anthropic.Anthropic()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

question = "Hello!"
with support.run(question) as run:  # one user message, one run
    answer = client.beta.messages.tool_runner(
        model="claude-haiku-5-5",
        max_tokens=4096,
        system=run.system,     # your prompt, then the skill list
        tools=[*run.tools],    # your tools, then load_skill
        messages=[{"role": "user", "content": question}],
    ).until_done()

Install

pip install "decimalai[anthropic]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import anthropic
from decimalai.anthropic import agent_config

client = anthropic.Anthropic()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

question = "Hello!"
with support.run(question) as run:  # one user message, one run
    answer = client.beta.messages.tool_runner(
        model="claude-haiku-5-5",
        max_tokens=4096,
        system=run.system,     # your prompt, then the skill list
        tools=[*run.tools],    # your tools, then load_skill
        messages=[{"role": "user", "content": question}],
    ).until_done()

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
import anthropic

client = anthropic.Anthropic()
skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
reply = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,
    system=f"You are a helpful agent.\n\n{skill}",
    messages=[{"role": "user", "content": "Hello!"}],
)

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Anthropic docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[google-genai]"
decimalai skills add openlair-moe-training

Add it to your agent

from google import genai
from google.genai import types
from decimalai.google_genai import agent_config

client = genai.Client()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")


def lookup_order(order_id: str) -> str:
    return f"Order {order_id} shipped yesterday."


# Your functions, as declarations for the model, and what answers each.
TOOLS = [{"function_declarations": [{
    "name": "lookup_order",
    "description": "Look up the status of an order.",
    "parameters_json_schema": {"type": "object", "properties": {"order_id": {"type": "string"}},
                               "required": ["order_id"]},
}]}]
FUNCTIONS = {"lookup_order": lookup_order}

question = "Where is order A1001?"
with support.run(question) as run:  # one user message, one run
    config = types.GenerateContentConfig(
        system_instruction=run.system,     # your prompt, then the skill list
        tools=[*TOOLS, *run.tool_params],  # your functions, then load_skill
    )
    contents = [{"role": "user", "parts": [{"text": run.message}]}]
    while True:
        response = client.models.generate_content(
            model="gemini-3.5-flash", contents=contents, config=config,
        )
        if not response.function_calls:
            break
        contents.append(response.candidates[0].content)
        parts = []
        for call in response.function_calls:
            # load_skill is answered by the run; your own functions by FUNCTIONS.
            # A loop of your own rather than automatic function calling, which
            # raises KeyError when the model asks for a function you did not pass.
            function = FUNCTIONS.get(call.name)
            parts.append(run.tool_result(call) or {"function_response": {
                "name": call.name, "id": call.id,
                "response": {"result": function(**(call.args or {})) if function
                             else f"There is no function named {call.name!r}."},
            }})
        contents.append({"role": "user", "parts": parts})
print(response.text)

Install

pip install "decimalai[google-genai]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from google import genai
from google.genai import types
from decimalai.google_genai import agent_config

client = genai.Client()
# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")


def lookup_order(order_id: str) -> str:
    return f"Order {order_id} shipped yesterday."


# Your functions, as declarations for the model, and what answers each.
TOOLS = [{"function_declarations": [{
    "name": "lookup_order",
    "description": "Look up the status of an order.",
    "parameters_json_schema": {"type": "object", "properties": {"order_id": {"type": "string"}},
                               "required": ["order_id"]},
}]}]
FUNCTIONS = {"lookup_order": lookup_order}

question = "Where is order A1001?"
with support.run(question) as run:  # one user message, one run
    config = types.GenerateContentConfig(
        system_instruction=run.system,     # your prompt, then the skill list
        tools=[*TOOLS, *run.tool_params],  # your functions, then load_skill
    )
    contents = [{"role": "user", "parts": [{"text": run.message}]}]
    while True:
        response = client.models.generate_content(
            model="gemini-3.5-flash", contents=contents, config=config,
        )
        if not response.function_calls:
            break
        contents.append(response.candidates[0].content)
        parts = []
        for call in response.function_calls:
            # load_skill is answered by the run; your own functions by FUNCTIONS.
            # A loop of your own rather than automatic function calling, which
            # raises KeyError when the model asks for a function you did not pass.
            function = FUNCTIONS.get(call.name)
            parts.append(run.tool_result(call) or {"function_response": {
                "name": call.name, "id": call.id,
                "response": {"result": function(**(call.args or {})) if function
                             else f"There is no function named {call.name!r}."},
            }})
        contents.append({"role": "user", "parts": parts})
print(response.text)

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from google import genai
from google.genai import types

client = genai.Client()
skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
response = client.models.generate_content(
    model="gemini-3.5-flash",
    contents="Hello!",
    config=types.GenerateContentConfig(system_instruction=f"You are a helpful agent.\n\n{skill}"),
)

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Gemini API docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[langchain]" langchain-openai
decimalai skills add openlair-moe-training

Add it to your agent

from langchain.agents import create_agent
from decimalai.langchain import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = create_agent(
    "openai:gpt-5.4-mini",
    tools=[],                         # your tools; the middleware adds load_skill
    middleware=[support.middleware],  # your prompt, the skill list and load_skill
)
result = agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

Install

pip install "decimalai[langchain]" langchain-openai

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from langchain.agents import create_agent
from decimalai.langchain import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = create_agent(
    "openai:gpt-5.4-mini",
    tools=[],                         # your tools; the middleware adds load_skill
    middleware=[support.middleware],  # your prompt, the skill list and load_skill
)
result = agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from langchain.agents import create_agent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = create_agent("openai:gpt-5.4-mini", tools=[], system_prompt=f"You are a helpful agent.\n\n{skill}")
result = agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

LangChain docs

CrewAI's own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" "crewai>=1.15.27"
decimalai skills add openlair-moe-training

Add it to your agent

import decimalai.crewai
from crewai import Agent, Crew, Task

agent = Agent(
    role="Support", goal="Help the customer", backstory="A support agent.",
    llm=llm,                               # your own CrewAI LLM
    skills=decimalai.crewai.skills(),      # the installed folder
    tools=decimalai.crewai.skill_tools(),  # reads a skill's files, runs its scripts
)
task = Task(description="Hello!", expected_output="an answer", agent=agent)
Crew(agents=[agent], tasks=[task]).kickoff()

The same install from Python, for a notebook or a setup script. CrewAI's own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" "crewai>=1.15.27"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import decimalai.crewai
from crewai import Agent, Crew, Task

agent = Agent(
    role="Support", goal="Help the customer", backstory="A support agent.",
    llm=llm,                               # your own CrewAI LLM
    skills=decimalai.crewai.skills(),      # the installed folder
    tools=decimalai.crewai.skill_tools(),  # reads a skill's files, runs its scripts
)
task = Task(description="Hello!", expected_output="an answer", agent=agent)
Crew(agents=[agent], tasks=[task]).kickoff()

No DecimalAI package. CrewAI's own loader reads the folder, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from crewai import Agent, Crew, Task

agent = Agent(
    role="Support", goal="Help the customer", backstory="A support agent.",
    llm=llm,                         # your own CrewAI LLM
    skills=[Path(".agents/skills")],
)
task = Task(description="Hello!", expected_output="an answer", agent=agent)
Crew(agents=[agent], tasks=[task]).kickoff()

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

CrewAI docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[adk]"
decimalai skills add openlair-moe-training

Add it to your agent

from decimalai.adk import DecimalAIPlugin, agent_config
from google.adk.agents import LlmAgent
from google.adk.apps import App

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = LlmAgent(
    name="support",          # ADK needs a Python identifier
    model="gemini-3.5-flash",
    instruction=support.instruction,  # your prompt, then the skill list
    tools=[*support.tools],           # your tools, then load_skill
)
# DecimalAIPlugin makes each invocation of runners built from this App one trace.
app = App(name="support", root_agent=agent, plugins=[DecimalAIPlugin()])

Prefer ADK's own loader? SkillToolset(skills=decimalai.adk.skills(), ...) reads the same folder. Use one or the other: a skill delivered both ways reaches the prompt twice.

Install

pip install "decimalai[adk]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from decimalai.adk import DecimalAIPlugin, agent_config
from google.adk.agents import LlmAgent
from google.adk.apps import App

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = LlmAgent(
    name="support",          # ADK needs a Python identifier
    model="gemini-3.5-flash",
    instruction=support.instruction,  # your prompt, then the skill list
    tools=[*support.tools],           # your tools, then load_skill
)
# DecimalAIPlugin makes each invocation of runners built from this App one trace.
app = App(name="support", root_agent=agent, plugins=[DecimalAIPlugin()])

Prefer ADK's own loader? SkillToolset(skills=decimalai.adk.skills(), ...) reads the same folder. Use one or the other: a skill delivered both ways reaches the prompt twice.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from google.adk.agents import LlmAgent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = LlmAgent(
    name="support",
    model="gemini-3.5-flash",
    instruction=lambda context: f"You are a helpful agent.\n\n{skill}",
)

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Google ADK docs

agent_config() hands your agent the skill. With no DecimalAI key it reads your project's files and sends nothing. Needs SDK 0.27 or newer.

Install

pip install "decimalai[pydantic-ai]"
decimalai skills add openlair-moe-training

Add it to your agent

from pydantic_ai import Agent
from decimalai.pydantic_ai import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = Agent(
    "openai:gpt-4o-mini",
    name=support.name,
    tools=[],                 # your tools; the capability adds load_skill
    capabilities=[support],   # your prompt, the skill list and load_skill
)
# In a notebook or async code: result = await agent.run("Hello!")
result = agent.run_sync("Hello!")

On the Pydantic AI harness, decimalai.pydantic_ai_harness.skills(project) uses its own loader instead.

Install

pip install "decimalai[pydantic-ai]"

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from pydantic_ai import Agent
from decimalai.pydantic_ai import agent_config

# With no DECIMAL_API_KEY this reads the skills your project lists, from disk.
support = agent_config("support", fallback="You are a helpful agent.")

agent = Agent(
    "openai:gpt-4o-mini",
    name=support.name,
    tools=[],                 # your tools; the capability adds load_skill
    capabilities=[support],   # your prompt, the skill list and load_skill
)
# In a notebook or async code: result = await agent.run("Hello!")
result = agent.run_sync("Hello!")

On the Pydantic AI harness, decimalai.pydantic_ai_harness.skills(project) uses its own loader instead.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from pydantic_ai import Agent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = Agent("openai:gpt-4o-mini", instructions=f"You are a helpful agent.\n\n{skill}")
# In a notebook or async code: result = await agent.run("Hello!")
result = agent.run_sync("Hello!")

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Pydantic AI docs

Deep Agents' own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" deepagents
decimalai skills add openlair-moe-training

Add it to your agent

import decimalai.deepagents
from deepagents import create_deep_agent

agent = create_deep_agent(model=model, **decimalai.deepagents.skills())
agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

The default backend reads files but has no shell. To run a skill's scripts, pass skills(backend=LocalShellBackend(...)).

The same install from Python, for a notebook or a setup script. Deep Agents' own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" deepagents

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import decimalai.deepagents
from deepagents import create_deep_agent

agent = create_deep_agent(model=model, **decimalai.deepagents.skills())
agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

The default backend reads files but has no shell. To run a skill's scripts, pass skills(backend=LocalShellBackend(...)).

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from deepagents import create_deep_agent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = create_deep_agent(model=model, system_prompt=f"You are a helpful agent.\n\n{skill}")
agent.invoke({"messages": [{"role": "user", "content": "Hello!"}]})

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Deep Agents docs

Its own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" agent-framework
decimalai skills add openlair-moe-training

Add it to your agent

import asyncio

from agent_framework import Agent
from decimalai.agent_framework import skills

agent = Agent(client=client, instructions="You are a helpful agent.", context_providers=[skills()])
asyncio.run(agent.run("Hello!"))

It caches skills by default, so restart the process after decimalai skills update.

The same install from Python, for a notebook or a setup script. Its own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" agent-framework

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

import asyncio

from agent_framework import Agent
from decimalai.agent_framework import skills

agent = Agent(client=client, instructions="You are a helpful agent.", context_providers=[skills()])
asyncio.run(agent.run("Hello!"))

It caches skills by default, so restart the process after decimalai skills update.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

import asyncio
from pathlib import Path
from agent_framework import Agent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = Agent(client=client, instructions=f"You are a helpful agent.\n\n{skill}")
asyncio.run(agent.run("Hello!"))

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Microsoft Agent Framework docs

Strands' own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" strands-agents
decimalai skills add openlair-moe-training

Add it to your agent

from decimalai.strands import skills
from strands import Agent
from strands.vended_tools import file_editor, shell

agent = Agent(
    model=model,                # always pass one: the default is Amazon Bedrock
    system_prompt="You are a helpful agent.",
    tools=[file_editor, shell],
    plugins=[skills()],
)
agent("Hello!")

Strands loads an agent's skills once, so build the agent again after decimalai skills update.

The same install from Python, for a notebook or a setup script. Strands' own skill loader reads the installed folder. No DecimalAI account needed.

Install

pip install "decimalai>=0.16" strands-agents

Add it to your agent

from decimalai import skillpkg

# Writes .agents/skills and .decimal/skills.toml. Safe to run twice.
result = skillpkg.add("openlair-moe-training")
if result.exit_code:
    raise SystemExit(result.problems)

from decimalai.strands import skills
from strands import Agent
from strands.vended_tools import file_editor, shell

agent = Agent(
    model=model,                # always pass one: the default is Amazon Bedrock
    system_prompt="You are a helpful agent.",
    tools=[file_editor, shell],
    plugins=[skills()],
)
agent("Hello!")

Strands loads an agent's skills once, so build the agent again after decimalai skills update.

No DecimalAI package. The whole skill goes in every prompt, and nothing pins or updates it.

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

Give it to your agent

from pathlib import Path
from strands import Agent

skill = Path(".agents/skills/openlair-moe-training/SKILL.md").read_text()
agent = Agent(model=model, system_prompt=f"You are a helpful agent.\n\n{skill}")
agent("Hello!")

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Strands docs

Adds every file the skill has, pinned so you can update it. Cursor, Copilot, Cline and Warp read this folder by themselves.

pip install "decimalai>=0.15"
decimalai skills add openlair-moe-training

No Python or account. Cursor, Copilot, Cline and Warp read this folder by themselves.

Save it into your project

curl -fsSL --create-dirs -o .agents/skills/openlair-moe-training/SKILL.md https://app.decimal.ai/s/openlair-moe-training/SKILL.md

Copy SKILL.md

Save it in your project, as

.agents/skills/openlair-moe-training/SKILL.md

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Cursor and other editors docs

Fetch the file and give its text to your model as instructions. No account or package needed.

Get the file

curl -fsSL https://app.decimal.ai/s/openlair-moe-training/SKILL.md

Read the summary as JSON

curl -fsSL https://app.decimal.ai/s/openlair-moe-training.json

This gets SKILL.md only. The skill's 3 other files are not included.

Other ways to add it

Other setups docs

You don't need a DecimalAI account.

Version: v1, tested 8/7/2026
Source: OpenLAIR/dr-claw ↗
License: NOASSERTION

Manage it in DecimalAI

Saved prompts and skills, pinned versions, and traces of this skill in your runs. DecimalAI does not run your agent.

Claude Code reads files from your project, so DecimalAI config does not apply there.

On Claude Agent SDK: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On OpenAI Agents: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On OpenAI SDK: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On Anthropic: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On Gemini API: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On LangChain: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On CrewAI: set DECIMAL_API_KEY and add decimalai.init(crewai=True) to record runs and skill use. The skill version stays in your repo.

On Google ADK: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On Pydantic AI: save the skill to an agent here, set DECIMAL_API_KEY, and the same code uses it with no command to run. Changes you save reach the next run with no redeploy.

On Deep Agents: set DECIMAL_API_KEY and add decimalai.init(langchain=True) to record runs and skill use. The skill version stays in your repo.

On Microsoft Agent Framework: set DECIMAL_API_KEY and add decimalai.init(otel=True, agent_name="support") to record runs and skill use. The skill version stays in your repo.

On Strands: set DECIMAL_API_KEY and add decimalai.init(otel=True, agent_name="support") to record runs and skill use. The skill version stays in your repo.

Editors read files from your project, so DecimalAI config does not apply there.

Anywhere else: record a use with decimalai.log_skill_activation() inside a trace.

Use this skill

SKILL.md

MoE Training: Mixture of Experts

When to Use This Skill

Use MoE Training when you need to:

Train larger models with limited compute (5× cost reduction vs dense models)
Scale model capacity without proportional compute increase
Achieve better performance per compute budget than dense models
Specialize experts for different domains/tasks/languages
Reduce inference latency with sparse activation (only 13B/47B params active in Mixtral)
Implement SOTA models like Mixtral 8x7B, DeepSeek-V3, Switch Transformers

Notable MoE Models: Mixtral 8x7B (Mistral AI), DeepSeek-V3, Switch Transformers (Google), GLaM (Google), NLLB-MoE (Meta)

Installation

# DeepSpeed with MoE support
pip install deepspeed>=0.6.0

# Megatron-DeepSpeed for large-scale training
git clone https://github.com/microsoft/Megatron-DeepSpeed
cd Megatron-DeepSpeed
pip install -r requirements.txt

# Alternative: HuggingFace Transformers
pip install transformers accelerate

Quick Start

Basic MoE Architecture

import torch
import torch.nn as nn

class MoELayer(nn.Module):
    """Sparse Mixture of Experts layer."""

    def __init__(self, hidden_size, num_experts=8, top_k=2):
        super().__init__()
        self.num_experts = num_experts
        self.top_k = top_k

        # Expert networks (FFN)
        self.experts = nn.ModuleList([
            nn.Sequential(
                nn.Linear(hidden_size, 4 * hidden_size),
                nn.GELU(),
                nn.Linear(4 * hidden_size, hidden_size)
            )
            for _ in range(num_experts)
        ])

        # Gating network (router)
        self.gate = nn.Linear(hidden_size, num_experts)

    def forward(self, x):
        # x shape: (batch_size, seq_len, hidden_size)
        batch_size, seq_len, hidden_size = x.shape

        # Flatten for routing
        x_flat = x.view(-1, hidden_size)  # (batch_size * seq_len, hidden_size)

        # Compute gate scores
        gate_logits = self.gate(x_flat)  # (batch_size * seq_len, num_experts)

        # Top-k routing
        gate_scores = torch.softmax(gate_logits, dim=-1)
        topk_scores, topk_indices = torch.topk(gate_scores, self.top_k, dim=-1)

        # Normalize top-k scores
        topk_scores = topk_scores / topk_scores.sum(dim=-1, keepdim=True)

        # Dispatch and combine expert outputs
        output = torch.zeros_like(x_flat)

        for i in range(self.top_k):
            expert_idx = topk_indices[:, i]
            expert_scores = topk_scores[:, i].unsqueeze(-1)

            # Route tokens to experts
            for expert_id in range(self.num_experts):
                mask = (expert_idx == expert_id)
                if mask.any():
                    expert_input = x_flat[mask]
                    expert_output = self.experts[expert_id](expert_input)
                    output[mask] += expert_scores[mask] * expert_output

        # Reshape back
        return output.view(batch_size, seq_len, hidden_size)

DeepSpeed MoE Training

# Training script with MoE
deepspeed pretrain_gpt_moe.py \
  --num-layers 24 \
  --hidden-size 1024 \
  --num-attention-heads 16 \
  --seq-length 2048 \
  --max-position-embeddings 2048 \
  --micro-batch-size 4 \
  --global-batch-size 256 \
  --train-iters 500000 \
  --lr 0.0001 \
  --min-lr 0.00001 \
  --lr-decay-style cosine \
  --num-experts 128 \
  --moe-expert-parallel-size 4 \
  --moe-loss-coeff 0.01 \
  --moe-train-capacity-factor 1.25 \
  --moe-eval-capacity-factor 2.0 \
  --fp16 \
  --deepspeed_config ds_config.json

Core Concepts

1. MoE Architecture

Key Components:

Experts: Multiple specialized FFN networks (typically 8-128)
Router/Gate: Learned network that selects which experts to use
Top-k Routing: Activate only k experts per token (k=1 or k=2)
Load Balancing: Ensure even expert utilization

Input Token
    ↓
Router (Gate Network)
    ↓
Top-k Expert Selection (e.g., 2 out of 8)
    ↓
Expert 1 (weight: 0.6) + Expert 5 (weight: 0.4)
    ↓
Weighted Combination
    ↓
Output

2. Routing Mechanisms

Top-1 Routing (Switch Transformer):

# Simplest routing: one expert per token
gate_logits = router(x)  # (batch, seq_len, num_experts)
expert_idx = torch.argmax(gate_logits, dim=-1)  # Hard routing

Top-2 Routing (Mixtral):

# Top-2: two experts per token
gate_scores = torch.softmax(router(x), dim=-1)
top2_scores, top2_indices = torch.topk(gate_scores, k=2, dim=-1)

# Normalize scores
top2_scores = top2_scores / top2_scores.sum(dim=-1, keepdim=True)

# Combine expert outputs
output = (top2_scores[:, :, 0:1] * expert_outputs[top2_indices[:, :, 0]] +
          top2_scores[:, :, 1:2] * expert_outputs[top2_indices[:, :, 1]])

Expert Choice Routing:

# Experts choose top-k tokens (instead of tokens choosing experts)
# Guarantees perfect load balancing
expert_scores = router(x).transpose(-1, -2)  # (batch, num_experts, seq_len)
topk_tokens = torch.topk(expert_scores, k=capacity_per_expert, dim=-1)

3. Load Balancing

Auxiliary Loss:

def load_balancing_loss(gate_logits, expert_indices, num_experts):
    """Encourage uniform expert usage."""
    # Fraction of tokens routed to each expert
    expert_counts = torch.bincount(expert_indices.flatten(), minlength=num_experts)
    expert_fraction = expert_counts.float() / expert_indices.numel()

    # Gate probability for each expert (average across tokens)
    gate_probs = torch.softmax(gate_logits, dim=-1).mean(dim=0)

    # Auxiliary loss: encourage alignment
    aux_loss = num_experts * (expert_fraction * gate_probs).sum()

    return aux_loss

# Add to main loss
total_loss = language_model_loss + 0.01 * load_balancing_loss(...)

Router Z-Loss (Stability):

def router_z_loss(logits):
    """Encourage router to have lower entropy (more decisive)."""
    z_loss = torch.logsumexp(logits, dim=-1).pow(2).mean()
    return z_loss

total_loss = lm_loss + 0.01 * aux_loss + 0.001 * router_z_loss(gate_logits)

4. Expert Parallelism

# DeepSpeed configuration
{
  "train_batch_size": 256,
  "fp16": {"enabled": true},
  "moe": {
    "enabled": true,
    "num_experts": 128,
    "expert_parallel_size": 8,  # Distribute 128 experts across 8 GPUs
    "capacity_factor": 1.25,    # Expert capacity = tokens_per_batch * capacity_factor / num_experts
    "drop_tokens": true,        # Drop tokens exceeding capacity
    "use_residual": false
  }
}

Training Configuration

DeepSpeed MoE Config

{
  "train_batch_size": 256,
  "gradient_accumulation_steps": 1,
  "optimizer": {
    "type": "Adam",
    "params": {
      "lr": 0.0001,
      "betas": [0.9, 0.999],
      "eps": 1e-8
    }
  },
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "initial_scale_power": 16
  },
  "moe": {
    "enabled": true,
    "num_experts": 128,
    "expert_parallel_size": 8,
    "moe_loss_coeff": 0.01,
    "train_capacity_factor": 1.25,
    "eval_capacity_factor": 2.0,
    "min_capacity": 4,
    "drop_tokens": true,
    "use_residual": false,
    "use_tutel": false
  },
  "zero_optimization": {
    "stage": 1
  }
}

Training Script

#!/bin/bash

# Mixtral-style MoE training
deepspeed --num_gpus 8 pretrain_moe.py \
  --model-parallel-size 1 \
  --num-layers 32 \
  --hidden-size 4096 \
  --num-attention-heads 32 \
  --seq-length 2048 \
  --max-position-embeddings 4096 \
  --micro-batch-size 2 \
  --global-batch-size 256 \
  --train-iters 500000 \
  --save-interval 5000 \
  --eval-interval 1000 \
  --eval-iters 100 \
  --lr 0.0001 \
  --min-lr 0.00001 \
  --lr-decay-style cosine \
  --lr-warmup-iters 2000 \
  --clip-grad 1.0 \
  --weight-decay 0.1 \
  --num-experts 8 \
  --moe-expert-parallel-size 4 \
  --moe-loss-coeff 0.01 \
  --moe-train-capacity-factor 1.25 \
  --moe-eval-capacity-factor 2.0 \
  --disable-moe-token-dropping \
  --fp16 \
  --deepspeed \
  --deepspeed_config ds_config_moe.json \
  --data-path /path/to/data \
  --vocab-file /path/to/vocab.json \
  --merge-file /path/to/merges.txt

Advanced Patterns

Mixtral 8x7B Architecture

class MixtralMoEBlock(nn.Module):
    """Mixtral-style MoE block with 8 experts, top-2 routing."""

    def __init__(self, config):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.ffn_dim = config.intermediate_size
        self.num_experts = config.num_local_experts  # 8
        self.top_k = config.num_experts_per_tok       # 2

        # 8 expert FFNs
        self.experts = nn.ModuleList([
            nn.Sequential(
                nn.Linear(self.hidden_dim, self.ffn_dim, bias=False),
                nn.SiLU(),
                nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
            )
            for _ in range(self.num_experts)
        ])

        # Router
        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)

    def forward(self, hidden_states):
        batch_size, sequence_length, hidden_dim = hidden_states.shape

        # Flatten
        hidden_states = hidden_states.view(-1, hidden_dim)

        # Router logits
        router_logits = self.gate(hidden_states)  # (batch * seq_len, num_experts)

        # Softmax and top-2
        routing_weights = torch.softmax(router_logits, dim=1)
        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)

        # Normalize routing weights
        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)

        # Initialize output
        final_hidden_states = torch.zeros_like(hidden_states)

        # Route to experts
        for expert_idx in range(self.num_experts):
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(selected_experts == expert_idx)

            if idx.shape[0] == 0:
                continue

            # Current expert tokens
            current_hidden_states = hidden_states[idx]

            # Expert forward
            current_hidden_states = expert_layer(current_hidden_states)

            # Weighted by routing scores
            current_hidden_states *= routing_weights[idx, top_x, None]

            # Accumulate
            final_hidden_states.index_add_(0, idx, current_hidden_states)

        # Reshape
        return final_hidden_states.view(batch_size, sequence_length, hidden_dim)

PR-MoE (Pyramid-Residual-MoE)

# DeepSpeed PR-MoE: 3x better parameter efficiency
deepspeed pretrain_gpt_moe.py \
  --num-layers 24 \
  --hidden-size 1024 \
  --num-attention-heads 16 \
  --num-experts "[128, 64, 32, 16]" \
  --mlp-type residual \
  --moe-expert-parallel-size 4 \
  --moe-loss-coeff 0.01 \
  --fp16

Best Practices

1. Expert Count Selection

# Rule of thumb: More experts = more capacity, but diminishing returns
# Typical configurations:
# - Small models (1B-7B): 8-16 experts
# - Medium models (7B-30B): 8-64 experts
# - Large models (30B+): 64-256 experts

# Example: Mixtral 8x7B
# Total params: 47B (8 experts × 7B each)
# Active params: 13B (2 experts × 7B, top-2 routing)
# Efficiency: 47B capacity with 13B compute

2. Capacity Factor Tuning

# Capacity = (tokens_per_batch / num_experts) * capacity_factor

# Training: Lower capacity (faster, drops some tokens)
train_capacity_factor = 1.25  # 25% buffer

# Evaluation: Higher capacity (no dropping)
eval_capacity_factor = 2.0    # 100% buffer

# Formula:
expert_capacity = int((seq_len * batch_size / num_experts) * capacity_factor)

3. Learning Rate Guidelines

# MoE models need lower LR than dense models
# - Dense model: lr = 6e-4
# - MoE model: lr = 1e-4 (3-6× lower)

# Also extend decay schedule
dense_lr_decay_iters = 300000
moe_lr_decay_iters = 500000  # 1.5-2× longer

4. Loss Coefficient Tuning

# Start with standard values
moe_loss_coeff = 0.01    # Auxiliary loss (load balancing)
router_z_loss_coeff = 0.001  # Router entropy (stability)

# If load imbalance persists, increase aux loss
if max_expert_usage / min_expert_usage > 2.0:
    moe_loss_coeff = 0.1  # Stronger load balancing

# If training unstable, increase z-loss
if grad_norm > 10.0:
    router_z_loss_coeff = 0.01

5. Avoid Common Pitfalls

# ❌ Bad: Using same LR as dense model
optimizer = Adam(model.parameters(), lr=6e-4)

# ✅ Good: Lower LR for MoE
optimizer = Adam([
    {'params': model.non_moe_params, 'lr': 6e-4},
    {'params': model.moe_params, 'lr': 1e-4}
])

# ❌ Bad: No load balancing
loss = lm_loss

# ✅ Good: Add auxiliary loss
loss = lm_loss + 0.01 * aux_loss + 0.001 * z_loss

# ❌ Bad: Too many experts for small dataset
num_experts = 128  # Overfitting risk

# ✅ Good: Match experts to data diversity
num_experts = 8  # Better for small datasets

Inference Optimization

Sparse Inference

# Only activate top-k experts (huge memory savings)
@torch.no_grad()
def moe_inference(x, model, top_k=2):
    """Sparse MoE inference: only load k experts."""
    # Router
    gate_logits = model.gate(x)
    topk_scores, topk_indices = torch.topk(
        torch.softmax(gate_logits, dim=-1),
        k=top_k,
        dim=-1
    )

    # Load and run only top-k experts
    output = torch.zeros_like(x)
    for i in range(top_k):
        expert_idx = topk_indices[:, i]
        # Load expert from disk/offload if needed
        expert = model.load_expert(expert_idx)
        output += topk_scores[:, i:i+1] * expert(x)

    return output

Resources

DeepSpeed MoE Tutorial: https://www.deepspeed.ai/tutorials/mixture-of-experts-nlg/
Mixtral Paper: https://arxiv.org/abs/2401.04088
Switch Transformers: https://arxiv.org/abs/2101.03961
HuggingFace MoE Guide: https://huggingface.co/blog/moe
NVIDIA MoE Blog: https://developer.nvidia.com/blog/applying-mixture-of-experts-in-llm-architectures/

Related skills

Skills that do similar work, with their headline benchmark lift.

From the same source and nearby in the catalogue.

Ready to supercharge your AI agents?

MoE Training: Mixture of Experts

SkillScore

Safety

What it changes

Track record

Use this skill

MoE Training: Mixture of Experts

When to Use This Skill

Installation

Quick Start

Basic MoE Architecture

DeepSpeed MoE Training

Core Concepts

1. MoE Architecture

2. Routing Mechanisms

3. Load Balancing

4. Expert Parallelism

Training Configuration

DeepSpeed MoE Config

Training Script

Advanced Patterns

Mixtral 8x7B Architecture

PR-MoE (Pyramid-Residual-MoE)

Best Practices

1. Expert Count Selection

2. Capacity Factor Tuning

3. Learning Rate Guidelines

4. Loss Coefficient Tuning

5. Avoid Common Pitfalls

Inference Optimization

Sparse Inference

Resources

See Also

When it activates

SkillSafety

Source

Related skills

Ready to supercharge your AI agents?

SkillScore

Safety

What it changes

Track record

Use this skill

MoE Training: Mixture of Experts

When to Use This Skill

Installation

Quick Start

Basic MoE Architecture

DeepSpeed MoE Training

Core Concepts

1. MoE Architecture

2. Routing Mechanisms

3. Load Balancing

4. Expert Parallelism

Training Configuration

DeepSpeed MoE Config

Training Script

Advanced Patterns

Mixtral 8x7B Architecture

PR-MoE (Pyramid-Residual-MoE)

Best Practices

1. Expert Count Selection

2. Capacity Factor Tuning

3. Learning Rate Guidelines

4. Loss Coefficient Tuning

5. Avoid Common Pitfalls

Inference Optimization

Sparse Inference

Resources

See Also

When it activates

SkillSafety

Source

Related skills

More in the registry