openenv/pr_837/en/_app/immutable/nodes/23.8ea12e89.js

import{s as Ft,n as Yt,o as Dt}from"../chunks/scheduler.2b22cead.js";import{S as Kt,i as qt,e as r,s as i,c as d,h as Ot,a,d as l,b as s,f as Wt,g as p,j as o,k as Pt,l as te,m as n,n as m,t as c,o as w,p as M}from"../chunks/index.1a0e8013.js";import{C as ee,H as J,E as le}from"../chunks/MermaidChart.svelte_svelte_type_style_lang.fdca7295.js";import{C as xt}from"../chunks/CodeBlock.44c9dc0c.js";function ne(gt){let u,Y,P,D,T,K,f,q,y,_t="RL environment for GPU kernel optimization. Train LLM agents to write fast CUDA/Triton kernels.",O,C,tt,b,St="Agents receive a PyTorch reference implementation and must write an optimized GPU kernel that:",et,U,Bt="<li>Produces the same output (within tolerance)</li> <li>Runs faster than the baseline</li>",lt,h,kt="Each submission is evaluated with:",nt,v,Gt="<li>Compilation checking</li> <li>Correctness verification against reference</li> <li>Benchmark timing for speedup measurement</li> <li>NSight Systems profiling (optional)</li> <li>NSight Compute profiling (optional)</li>",it,$,st,I,rt,j,Qt="Requires: NVIDIA GPU with CUDA toolkit, PyTorch, Triton",at,x,ot,g,dt,_,pt,S,mt,B,ct,k,Zt="<thead><tr><th>Level</th> <th>Name</th> <th>Count</th> <th>Description</th></tr></thead> <tbody><tr><td>1</td> <td>Simple Operators</td> <td>15</td> <td>matmul, softmax, conv, norms</td></tr> <tr><td>2</td> <td>Fused Operations</td> <td>15</td> <td>matmul+activation chains</td></tr> <tr><td>3</td> <td>Single Blocks</td> <td>3</td> <td>attention, transformer block</td></tr> <tr><td>4</td> <td>Novel Layers</td> <td>8</td> <td>MLA, MoE, GQA, FP8, INT4</td></tr> <tr><td>5</td> <td>Scientific Computing</td> <td>8</td> <td>N-body, stencil, SpMV</td></tr> <tr><td>6</td> <td>Graphics</td> <td>8</td> <td>ray tracing, histogram, blur</td></tr> <tr><td>7</td> <td>Signal Processing</td> <td>8</td> <td>FFT, convolution, median filter</td></tr> <tr><td>8</td> <td>Video Processing</td> <td>8</td> <td>motion estimation, optical flow</td></tr> <tr><td>9</td> <td>Parallel Primitives</td> <td>8</td> <td>scan, reduction, radix sort</td></tr> <tr><td>10</td> <td>Cryptography</td> <td>8</td> <td>SHA-256, AES, ChaCha20</td></tr></tbody>",wt,G,Lt="<strong>Total: 89 problems</strong>",Mt,Q,ut,Z,Ht=`Rewards are designed so that <strong>only speedup &gt; 1.0x baseline produces positive reward</strong>.
Compilation and correctness alone do not give positive reward - they are necessary but not sufficient.`,Jt,L,Et="<thead><tr><th>Condition</th> <th>Reward</th> <th>Description</th></tr></thead> <tbody><tr><td>Compilation failure</td> <td>-0.5</td> <td>Penalty for code that doesn’t compile</td></tr> <tr><td>Correctness failure</td> <td>-0.25</td> <td>Penalty for incorrect output</td></tr> <tr><td>Correct but slower</td> <td>(speedup - 1.0) * 0.5</td> <td>Small negative for being slower than baseline</td></tr> <tr><td>Correct and faster</td> <td>min(speedup - 1.0, 2.0)</td> <td>Positive, capped at 2.0</td></tr></tbody>",Tt,H,Rt="<strong>Examples:</strong>",ft,E,At="<li>Compile fail: reward = -0.5</li> <li>Compiles, wrong output: reward = -0.25</li> <li>Compiles, correct, 0.8x speed: reward = -0.1</li> <li>Compiles, correct, 1.0x speed: reward = 0.0</li> <li>Compiles, correct, 1.5x speed: reward = 0.5</li> <li>Compiles, correct, 3.0x speed: reward = 2.0 (capped)</li>",yt,R,Ct,A,Xt=`<strong>Warning:</strong> This environment executes user-submitted kernel code with full Python/CUDA privileges.
While Docker provides container isolation, there is no sandboxing within the container for:`,bt,X,zt="<li>Filesystem access</li> <li>Network requests</li> <li>Resource consumption (GPU memory, CPU)</li> <li>Module imports</li>",Ut,z,Nt=`This is acceptable for trusted research environments but should be documented as a security consideration.
For production deployments, consider additional isolation measures.`,ht,N,vt,V,Vt="BSD-3-Clause (following OpenEnv licensing)",$t,W,It,F,jt;return T=new ee({props:{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"}}),f=new J({props:{title:"kernrl",local:"kernrl",headingTag:"h1"}}),C=new J({props:{title:"Overview",local:"overview",headingTag:"h2"}}),$=new J({props:{title:"Installation",local:"installation",headingTag:"h2"}}),I=new xt({props:{code:"Y2QlMjBlbnZzJTJGa2VybnJsJTBBcGlwJTIwaW5zdGFsbCUyMC1lJTIwLg==",highlighted:`<span class="hljs-built_in">cd</span> envs/kernrl
pip install -e .`,lang:"bash",wrap:!1}}),x=new J({props:{title:"Quick Start",local:"quick-start",headingTag:"h2"}}),g=new xt({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> kernrl <span class="hljs-keyword">import</span> KernelAction, kernrl_env

<span class="hljs-comment"># Connect to server</span>
env = kernrl_env(base_url=<span class="hljs-string">&quot;http://localhost:8000&quot;</span>)

<span class="hljs-comment"># Start episode</span>
obs = env.reset(problem_id=<span class="hljs-string">&quot;L1_23_Softmax&quot;</span>)
<span class="hljs-built_in">print</span>(obs.problem_description)

<span class="hljs-comment"># Submit a kernel</span>
action = KernelAction(code=<span class="hljs-string">&#x27;&#x27;&#x27;
import torch
import triton
import triton.language as tl

@triton.jit
def softmax_kernel(input_ptr, output_ptr, n_cols, BLOCK_SIZE: tl.constexpr):
    row_idx = tl.program_id(0)
    col_offsets = tl.arange(0, BLOCK_SIZE)
    mask = col_offsets &lt; n_cols

    row_start = row_idx * n_cols
    row = tl.load(input_ptr + row_start + col_offsets, mask=mask, other=-float(&#x27;inf&#x27;))

    row_max = tl.max(row, axis=0)
    row = row - row_max
    numerator = tl.exp(row)
    denominator = tl.sum(numerator, axis=0)
    softmax_output = numerator / denominator

    tl.store(output_ptr + row_start + col_offsets, softmax_output, mask=mask)

class Model(torch.nn.Module):
    def forward(self, x):
        n_rows, n_cols = x.shape
        output = torch.empty_like(x)
        BLOCK_SIZE = triton.next_power_of_2(n_cols)
        softmax_kernel[(n_rows,)](x, output, n_cols, BLOCK_SIZE=BLOCK_SIZE)
        return output
&#x27;&#x27;&#x27;</span>)

result = env.step(action)
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Speedup: <span class="hljs-subst">{result.observation.speedup}</span>x&quot;</span>)
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Correct: <span class="hljs-subst">{result.observation.correctness_pass}</span>&quot;</span>)`,lang:"python",wrap:!1}}),_=new J({props:{title:"Running the Server",local:"running-the-server",headingTag:"h2"}}),S=new xt({props:{code:"JTIzJTIwRGV2ZWxvcG1lbnQlMEF1dmljb3JuJTIwa2VybnJsLnNlcnZlci5hcHAlM0FhcHAlMjAtLXJlbG9hZCUyMC0taG9zdCUyMDAuMC4wLjAlMjAtLXBvcnQlMjA4MDAwJTBBJTBBJTIzJTIwRG9ja2VyJTIwKEdQVSUyMHJlcXVpcmVkKSUwQWNkJTIwZW52cyUyRmtlcm5ybCUwQWRvY2tlciUyMGJ1aWxkJTIwLXQlMjBrZXJucmwlMjAtZiUyMHNlcnZlciUyRkRvY2tlcmZpbGUlMjAuJTBBZG9ja2VyJTIwcnVuJTIwLS1ncHVzJTIwYWxsJTIwLXAlMjA4MDAwJTNBODAwMCUyMGtlcm5ybA==",highlighted:`<span class="hljs-comment"># Development</span>
uvicorn kernrl.server.app:app --reload --host 0.0.0.0 --port 8000

<span class="hljs-comment"># Docker (GPU required)</span>
<span class="hljs-built_in">cd</span> envs/kernrl
docker build -t kernrl -f server/Dockerfile .
docker run --gpus all -p 8000:8000 kernrl`,lang:"bash",wrap:!1}}),B=new J({props:{title:"Problem Levels",local:"problem-levels",headingTag:"h2"}}),Q=new J({props:{title:"Reward Structure",local:"reward-structure",headingTag:"h2"}}),R=new J({props:{title:"Security Considerations",local:"security-considerations",headingTag:"h2"}}),N=new J({props:{title:"License",local:"license",headingTag:"h2"}}),W=new le({props:{source:"https://github.com/huggingface/openenv/blob/main/docs/source/environments/kernrl.md"}}),{c(){u=r("meta"),Y=i(),P=r("p"),D=i(),d(T.$$.fragment),K=i(),d(f.$$.fragment),q=i(),y=r("p"),y.textContent=_t,O=i(),d(C.$$.fragment),tt=i(),b=r("p"),b.textContent=St,et=i(),U=r("ol"),U.innerHTML=Bt,lt=i(),h=r("p"),h.textContent=kt,nt=i(),v=r("ul"),v.innerHTML=Gt,it=i(),d($.$$.fragment),st=i(),d(I.$$.fragment),rt=i(),j=r("p"),j.textContent=Qt,at=i(),d(x.$$.fragment),ot=i(),d(g.$$.fragment),dt=i(),d(_.$$.fragment),pt=i(),d(S.$$.fragment),mt=i(),d(B.$$.fragment),ct=i(),k=r("table"),k.innerHTML=Zt,wt=i(),G=r("p"),G.innerHTML=Lt,Mt=i(),d(Q.$$.fragment),ut=i(),Z=r("p"),Z.innerHTML=Ht,Jt=i(),L=r("table"),L.innerHTML=Et,Tt=i(),H=r("p"),H.innerHTML=Rt,ft=i(),E=r("ul"),E.innerHTML=At,yt=i(),d(R.$$.fragment),Ct=i(),A=r("p"),A.innerHTML=Xt,bt=i(),X=r("ul"),X.innerHTML=zt,Ut=i(),z=r("p"),z.textContent=Nt,ht=i(),d(N.$$.fragment),vt=i(),V=r("p"),V.textContent=Vt,$t=i(),d(W.$$.fragment),It=i(),F=r("p"),this.h()},l(t){const e=Ot("svelte-u9bgzb",document.head);u=a(e,"META",{name:!0,content:!0}),e.forEach(l),Y=s(t),P=a(t,"P",{}),Wt(P).forEach(l),D=s(t),p(T.$$.fragment,t),K=s(t),p(f.$$.fragment,t),q=s(t),y=a(t,"P",{"data-svelte-h":!0}),o(y)!=="svelte-1ny0dk7"&&(y.textContent=_t),O=s(t),p(C.$$.fragment,t),tt=s(t),b=a(t,"P",{"data-svelte-h":!0}),o(b)!=="svelte-qjdxt7"&&(b.textContent=St),et=s(t),U=a(t,"OL",{"data-svelte-h":!0}),o(U)!=="svelte-1rl5hha"&&(U.innerHTML=Bt),lt=s(t),h=a(t,"P",{"data-svelte-h":!0}),o(h)!=="svelte-1m0plru"&&(h.textContent=kt),nt=s(t),v=a(t,"UL",{"data-svelte-h":!0}),o(v)!=="svelte-tnmxo9"&&(v.innerHTML=Gt),it=s(t),p($.$$.fragment,t),st=s(t),p(I.$$.fragment,t),rt=s(t),j=a(t,"P",{"data-svelte-h":!0}),o(j)!=="svelte-1kp177"&&(j.textContent=Qt),at=s(t),p(x.$$.fragment,t),ot=s(t),p(g.$$.fragment,t),dt=s(t),p(_.$$.fragment,t),pt=s(t),p(S.$$.fragment,t),mt=s(t),p(B.$$.fragment,t),ct=s(t),k=a(t,"TABLE",{"data-svelte-h":!0}),o(k)!=="svelte-qhgs1p"&&(k.innerHTML=Zt),wt=s(t),G=a(t,"P",{"data-svelte-h":!0}),o(G)!=="svelte-ic7naq"&&(G.innerHTML=Lt),Mt=s(t),p(Q.$$.fragment,t),ut=s(t),Z=a(t,"P",{"data-svelte-h":!0}),o(Z)!=="svelte-1vgwv0b"&&(Z.innerHTML=Ht),Jt=s(t),L=a(t,"TABLE",{"data-svelte-h":!0}),o(L)!=="svelte-1r66vsq"&&(L.innerHTML=Et),Tt=s(t),H=a(t,"P",{"data-svelte-h":!0}),o(H)!=="svelte-1r5zqy"&&(H.innerHTML=Rt),ft=s(t),E=a(t,"UL",{"data-svelte-h":!0}),o(E)!=="svelte-1unj6t4"&&(E.innerHTML=At),yt=s(t),p(R.$$.fragment,t),Ct=s(t),A=a(t,"P",{"data-svelte-h":!0}),o(A)!=="svelte-1tljewc"&&(A.innerHTML=Xt),bt=s(t),X=a(t,"UL",{"data-svelte-h":!0}),o(X)!=="svelte-11wrlfa"&&(X.innerHTML=zt),Ut=s(t),z=a(t,"P",{"data-svelte-h":!0}),o(z)!=="svelte-18sjlhc"&&(z.textContent=Nt),ht=s(t),p(N.$$.fragment,t),vt=s(t),V=a(t,"P",{"data-svelte-h":!0}),o(V)!=="svelte-1pqtjqe"&&(V.textContent=Vt),$t=s(t),p(W.$$.fragment,t),It=s(t),F=a(t,"P",{}),Wt(F).forEach(l),this.h()},h(){Pt(u,"name","hf:doc:metadata"),Pt(u,"content",ie)},m(t,e){te(document.head,u),n(t,Y,e),n(t,P,e),n(t,D,e),m(T,t,e),n(t,K,e),m(f,t,e),n(t,q,e),n(t,y,e),n(t,O,e),m(C,t,e),n(t,tt,e),n(t,b,e),n(t,et,e),n(t,U,e),n(t,lt,e),n(t,h,e),n(t,nt,e),n(t,v,e),n(t,it,e),m($,t,e),n(t,st,e),m(I,t,e),n(t,rt,e),n(t,j,e),n(t,at,e),m(x,t,e),n(t,ot,e),m(g,t,e),n(t,dt,e),m(_,t,e),n(t,pt,e),m(S,t,e),n(t,mt,e),m(B,t,e),n(t,ct,e),n(t,k,e),n(t,wt,e),n(t,G,e),n(t,Mt,e),m(Q,t,e),n(t,ut,e),n(t,Z,e),n(t,Jt,e),n(t,L,e),n(t,Tt,e),n(t,H,e),n(t,ft,e),n(t,E,e),n(t,yt,e),m(R,t,e),n(t,Ct,e),n(t,A,e),n(t,bt,e),n(t,X,e),n(t,Ut,e),n(t,z,e),n(t,ht,e),m(N,t,e),n(t,vt,e),n(t,V,e),n(t,$t,e),m(W,t,e),n(t,It,e),n(t,F,e),jt=!0},p:Yt,i(t){jt||(c(T.$$.fragment,t),c(f.$$.fragment,t),c(C.$$.fragment,t),c($.$$.fragment,t),c(I.$$.fragment,t),c(x.$$.fragment,t),c(g.$$.fragment,t),c(_.$$.fragment,t),c(S.$$.fragment,t),c(B.$$.fragment,t),c(Q.$$.fragment,t),c(R.$$.fragment,t),c(N.$$.fragment,t),c(W.$$.fragment,t),jt=!0)},o(t){w(T.$$.fragment,t),w(f.$$.fragment,t),w(C.$$.fragment,t),w($.$$.fragment,t),w(I.$$.fragment,t),w(x.$$.fragment,t),w(g.$$.fragment,t),w(_.$$.fragment,t),w(S.$$.fragment,t),w(B.$$.fragment,t),w(Q.$$.fragment,t),w(R.$$.fragment,t),w(N.$$.fragment,t),w(W.$$.fragment,t),jt=!1},d(t){t&&(l(Y),l(P),l(D),l(K),l(q),l(y),l(O),l(tt),l(b),l(et),l(U),l(lt),l(h),l(nt),l(v),l(it),l(st),l(rt),l(j),l(at),l(ot),l(dt),l(pt),l(mt),l(ct),l(k),l(wt),l(G),l(Mt),l(ut),l(Z),l(Jt),l(L),l(Tt),l(H),l(ft),l(E),l(yt),l(Ct),l(A),l(bt),l(X),l(Ut),l(z),l(ht),l(vt),l(V),l($t),l(It),l(F)),l(u),M(T,t),M(f,t),M(C,t),M($,t),M(I,t),M(x,t),M(g,t),M(_,t),M(S,t),M(B,t),M(Q,t),M(R,t),M(N,t),M(W,t)}}}const ie='{"title":"kernrl","local":"kernrl","sections":[{"title":"Overview","local":"overview","sections":[],"depth":2},{"title":"Installation","local":"installation","sections":[],"depth":2},{"title":"Quick Start","local":"quick-start","sections":[],"depth":2},{"title":"Running the Server","local":"running-the-server","sections":[],"depth":2},{"title":"Problem Levels","local":"problem-levels","sections":[],"depth":2},{"title":"Reward Structure","local":"reward-structure","sections":[],"depth":2},{"title":"Security Considerations","local":"security-considerations","sections":[],"depth":2},{"title":"License","local":"license","sections":[],"depth":2}],"depth":1}';function se(gt){return Dt(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class pe extends Kt{constructor(u){super(),qt(this,u,se,ne,Ft,{})}}export{pe as component};