Benchmarking DeepSeek R1 - 16x3090 #266
Replies: 5 comments 4 replies
-
Mixed quant of
|
| Component | Blocks 0-2 | Blocks 3-8 | Blocks 9-60 |
|---|---|---|---|
| Attention Query/Key/Value | q8_0 | q8_0 | q8_0 |
| Attention Output | q8_0 | q8_0 | q8_0 |
| FFN Down (regular) | q8_0 | - | - |
| FFN Gate/Up (regular) | q8_0 | - | - |
| FFN Down Shared Experts | - | q5_K | q5_K |
| FFN Gate/Up Shared Experts | - | q5_K | q5_K |
| FFN Down Experts | - | q5_K | iq4_xs |
| FFN Gate/Up Experts | - | iq4_xs | iq3_s |
| Output Layer | q8_0 | q8_0 | q8_0 |
| Compression Results | |||
| Original size: 1,282,038 MB (~1.2 TB) | |||
| Quantized size: 314,569 MB (~307 GB) | |||
| Compression ratio: 4.1x |
PPL
perplexity: tokenizing the input ..
perplexity: tokenization took 1195.26 ms
perplexity: calculating perplexity over 561 chunks, n_ctx=512, batch_size=2048, n_seq=4
perplexity: 11.69 seconds per pass - ETA 27.32 minutes
[1]2.5779,[2]3.3447,[3]2.4073,[4]2.0140,[5]1.8352,[6]1.6862,[7]1.5895,[8]1.5208,[9]1.4715,[10]1.4284,[11]1.4147,[12]1.4406,[13]1.4529,[14]1.5824,[15]1.7144,[16]1.7752,[17]1.9408,[18]2.0703,[19]2.0333,[20]2.0250,[21]2.1305,[22]2.1021,[23]2.0764,[24]2.0880,[25]2.0581,[26]2.0330,[27]2.0797,[28]2.0888,[29]2.1391,[30]2.1698,[31]2.2044,[32]2.2227,[33]2.2626,[34]2.3049,[35]2.3566,[36]2.4115,[37]2.4463,[38]2.4930,[39]2.5346,[40]2.5926,[41]2.6353,[42]2.6458,[43]2.6948,[44]2.7107,[45]2.7909,[46]2.8420,[47]2.8003,[48]2.7549,[49]2.7298,[50]2.7498,[51]2.7964,[52]2.8105,[53]2.8597,[54]2.8734,[55]2.9047,[56]2.9384,[57]2.9550,[58]2.9926,[59]3.0027,[60]3.0502,[61]3.0906,[62]3.1475,[63]3.1812,[64]3.2262,[65]3.2360,[66]3.2179,[67]3.1954,[68]3.2271,[69]3.2225,[70]3.2377,[71]3.2562,[72]3.2726,[73]3.2860,[74]3.3095,[75]3.2881,[76]3.2396,[77]3.1959,[78]3.1931,[79]3.1728,[80]3.1563,[81]3.1190,[82]3.1220,[83]3.0918,[84]3.0554,[85]3.0218,[86]2.9995,[87]2.9958,[88]2.9686,[89]2.9537,[90]2.9261,[91]2.8966,[92]2.8704,[93]2.8441,[94]2.8196,[95]2.7964,[96]2.7947,[97]2.8024,[98]2.7882,[99]2.7728,[100]2.7752,[101]2.7671,[102]2.7843,[103]2.8105,[104]2.8288,[105]2.8261,[106]2.8486,[107]2.8737,[108]2.8953,[109]2.9296,[110]2.9637,[111]2.9837,[112]2.9567,[113]2.9436,[114]2.9207,[115]2.9047,[116]2.8905,[117]2.8672,[118]2.8450,[119]2.8235,[120]2.8040,[121]2.7884,[122]2.7698,[123]2.7532,[124]2.7334,[125]2.7156,[126]2.6981,[127]2.6840,[128]2.6757,[129]2.6662,[130]2.6551,[131]2.6472,[132]2.6548,[133]2.6649,[134]2.6714,[135]2.6822,[136]2.6990,[137]2.7145,[138]2.7231,[139]2.7348,[140]2.7353,[141]2.7368,[142]2.7356,[143]2.7359,[144]2.7320,[145]2.7228,[146]2.7211,[147]2.7254,[148]2.7248,[149]2.7265,[150]2.7210,[151]2.7192,[152]2.7157,[153]2.7114,[154]2.7119,[155]2.7159,[156]2.7180,[157]2.7237,[158]2.7322,[159]2.7339,[160]2.7428,[161]2.7509,[162]2.7605,[163]2.7660,[164]2.7863,[165]2.8095,[166]2.8270,[167]2.8399,[168]2.8647,[169]2.8872,[170]2.9083,[171]2.9311,[172]2.9150,[173]2.8980,[174]2.8843,[175]2.8712,[176]2.8589,[177]2.8467,[178]2.8338,[179]2.8193,[180]2.8228,[181]2.8370,[182]2.8519,[183]2.8669,[184]2.8813,[185]2.8915,[186]2.9083,[187]2.9241,[188]2.9381,[189]2.9489,[190]2.9490,[191]2.9561,[192]2.9601,[193]2.9652,[194]2.9848,[195]2.9935,[196]3.0068,[197]3.0167,[198]3.0211,[199]3.0267,[200]3.0261,[201]3.0415,[202]3.0361,[203]3.0413,[204]3.0446,[205]3.0447,[206]3.0468,[207]3.0552,[208]3.0645,[209]3.0737,[210]3.0738,[211]3.0688,[212]3.0689,[213]3.0765,[214]3.0781,[215]3.0837,[216]3.0847,[217]3.0805,[218]3.0804,[219]3.0811,[220]3.0800,[221]3.0803,[222]3.0803,[223]3.0805,[224]3.0856,[225]3.0871,[226]3.0791,[227]3.0772,[228]3.0792,[229]3.0835,[230]3.0900,[231]3.0962,[232]3.0880,[233]3.0801,[234]3.0803,[235]3.0787,[236]3.0879,[237]3.0957,[238]3.1050,[239]3.1151,[240]3.1241,[241]3.1353,[242]3.1498,[243]3.1632,[244]3.1713,[245]3.1831,[246]3.1937,[247]3.1927,[248]3.1884,[249]3.1867,[250]3.1804,[251]3.1782,[252]3.1805,[253]3.1841,[254]3.1910,[255]3.1971,[256]3.2005,[257]3.2032,[258]3.2042,[259]3.2076,[260]3.2098,[261]3.2107,[262]3.2099,[263]3.2158,[264]3.2179,[265]3.2182,[266]3.2199,[267]3.2230,[268]3.2267,[269]3.2298,[270]3.2290,[271]3.2271,[272]3.2205,[273]3.2208,[274]3.2143,[275]3.2037,[276]3.1934,[277]3.1951,[278]3.2052,[279]3.2115,[280]3.2195,[281]3.2272,[282]3.2333,[283]3.2398,[284]3.2466,[285]3.2603,[286]3.2626,[287]3.2661,[288]3.2707,[289]3.2732,[290]3.2648,[291]3.2557,[292]3.2544,[293]3.2536,[294]3.2513,[295]3.2487,[296]3.2507,[297]3.2513,[298]3.2562,[299]3.2620,[300]3.2651,[301]3.2691,[302]3.2713,[303]3.2734,[304]3.2726,[305]3.2845,[306]3.2922,[307]3.3033,[308]3.2916,[309]3.2865,[310]3.2769,[311]3.2804,[312]3.2825,[313]3.2893,[314]3.2915,[315]3.2946,[316]3.2959,[317]3.2974,[318]3.2979,[319]3.2982,[320]3.3026,[321]3.3028,[322]3.3042,[323]3.3106,[324]3.3112,[325]3.3167,[326]3.3214,[327]3.3255,[328]3.3282,[329]3.3297,[330]3.3360,[331]3.3396,[332]3.3443,[333]3.3428,[334]3.3425,[335]3.3428,[336]3.3429,[337]3.3437,[338]3.3441,[339]3.3466,[340]3.3502,[341]3.3555,[342]3.3649,[343]3.3744,[344]3.3797,[345]3.3713,[346]3.3640,[347]3.3597,[348]3.3523,[349]3.3488,[350]3.3471,[351]3.3521,[352]3.3671,[353]3.3761,[354]3.3892,[355]3.3977,[356]3.4029,[357]3.4148,[358]3.4246,[359]3.4279,[360]3.4346,[361]3.4439,[362]3.4526,[363]3.4586,[364]3.4649,[365]3.4715,[366]3.4822,[367]3.4909,[368]3.4975,[369]3.5054,[370]3.5138,[371]3.5277,[372]3.5368,[373]3.5401,[374]3.5435,[375]3.5485,[376]3.5616,[377]3.5727,[378]3.5754,[379]3.5749,[380]3.5715,[381]3.5762,[382]3.5816,[383]3.5853,[384]3.5894,[385]3.5931,[386]3.5996,[387]3.6055,[388]3.6087,[389]3.5980,[390]3.5883,[391]3.5774,[392]3.5715,[393]3.5623,[394]3.5535,[395]3.5438,[396]3.5336,[397]3.5245,[398]3.5146,[399]3.5042,[400]3.4963,[401]3.4863,[402]3.4756,[403]3.4668,[404]3.4563,[405]3.4465,[406]3.4364,[407]3.4270,[408]3.4178,[409]3.4090,[410]3.4031,[411]3.4038,[412]3.3993,[413]3.4012,[414]3.4038,[415]3.4009,[416]3.4009,[417]3.4034,[418]3.3979,[419]3.3991,[420]3.3966,[421]3.3953,[422]3.3970,[423]3.3964,[424]3.4006,[425]3.4005,[426]3.4009,[427]3.3997,[428]3.4021,[429]3.4037,[430]3.4064,[431]3.4074,[432]3.4064,[433]3.4027,[434]3.4028,[435]3.3956,[436]3.3891,[437]3.3851,[438]3.3833,[439]3.3805,[440]3.3855,[441]3.3905,[442]3.3979,[443]3.3964,[444]3.3972,[445]3.3983,[446]3.4029,[447]3.4058,[448]3.4083,[449]3.4114,[450]3.4154,[451]3.4184,[452]3.4206,[453]3.4223,[454]3.4208,[455]3.4229,[456]3.4232,[457]3.4257,[458]3.4311,[459]3.4317,[460]3.4318,[461]3.4284,[462]3.4322,[463]3.4396,[464]3.4448,[465]3.4381,[466]3.4361,[467]3.4344,[468]3.4355,[469]3.4328,[470]3.4301,[471]3.4304,[472]3.4311,[473]3.4304,[474]3.4295,[475]3.4308,[476]3.4290,[477]3.4282,[478]3.4288,[479]3.4307,[480]3.4334,[481]3.4290,[482]3.4325,[483]3.4316,[484]3.4353,[485]3.4416,[486]3.4444,[487]3.4479,[488]3.4531,[489]3.4555,[490]3.4603,[491]3.4665,[492]3.4709,[493]3.4707,[494]3.4719,[495]3.4746,[496]3.4764,[497]3.4794,[498]3.4798,[499]3.4790,[500]3.4832,[501]3.4877,[502]3.4865,[503]3.4849,[504]3.4871,[505]3.4905,[506]3.4988,[507]3.5016,[508]3.5050,[509]3.4973,[510]3.4914,[511]3.4851,[512]3.4810,[513]3.4750,[514]3.4738,[515]3.4761,[516]3.4714,[517]3.4713,[518]3.4704,[519]3.4710,[520]3.4755,[521]3.4744,[522]3.4730,[523]3.4790,[524]3.4775,[525]3.4761,[526]3.4715,[527]3.4663,[528]3.4628,[529]3.4599,[530]3.4568,[531]3.4536,[532]3.4479,[533]3.4415,[534]3.4370,[535]3.4382,[536]3.4410,[537]3.4443,[538]3.4469,[539]3.4496,[540]3.4550,[541]3.4584,[542]3.4607,[543]3.4552,[544]3.4512,[545]3.4508,[546]3.4440,[547]3.4374,[548]3.4307,[549]3.4240,[550]3.4178,[551]3.4116,[552]3.4060,[553]3.4002,[554]3.3983,[555]3.3970,[556]3.3998,[557]3.4039,[558]3.4098,[559]3.4145,[560]3.4197,[561]3.4178,
Final estimate: PPL = 3.4178 +/- 0.01891
Beta Was this translation helpful? Give feedback.
-
|
Were you thinking of uploading this to huggingface, by any chance? I can reproduce and upload it myself if necessary, but I haven't downloaded the full R1 weights yet, and would be happy to continue avoiding that if possible! |
Beta Was this translation helpful? Give feedback.
-
|
@fredlas do you have any specific hardware configuration in mind? e.g. how much system RAM, and GPUs / VRAM? I put together rough notes on making your own custom quant in this quick-start guide discussion. I believe @davidsyoung has tailored the quant specific to his 16x3090 = 384 GB VRAM setup. I've made a couple quants now and have one okay one for 256GB RAM + 24GB VRAM single GPU configuration with better perplexity than unsloth You can get started with |
Beta Was this translation helpful? Give feedback.
-
|
I might be able to upload if you give me enough time, however, I actually recommend getting used to quanting as there’s a lot tweaking you may want to do. For example, I don’t actually think this quant suits my setup best yet, and I’m actually underutilising one GPU. I just haven’t found a way to split the layers that well yet. |
Beta Was this translation helpful? Give feedback.
-
|
@ubergarm 307GiB happens to be right around the size I'm thinking of. 72GiB VRAM + 256GiB RAM, for queuing up jobs to run overnight with 16k context - should just fit in there, I think. Funny coincidence for an extremely different configuration! Thanks for that guide - I made my own quants of Wizard2 8x22B a while back, but long enough that I was probably going to have to basically relearn it. @davidsyoung I'd say don't upload them just for my sake if you weren't already planning to - I just thought I'd check in case I could stay lazy. Plus this size range is probably pretty niche anyways; might not really be worth it in terms of helping people. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for this. I think it can be really useful for people. |
Beta Was this translation helpful? Give feedback.
-
|
@ikawrakow Can I convert this to a discussion? |
Beta Was this translation helpful? Give feedback.
-
|
All good with me @saood06 |
Beta Was this translation helpful? Give feedback.
-
Sure, go ahead |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Wanted to create a resource for anyone looking to optimise
-b -ub -ambwith-mla 2 -fa -fmoewith offloading DeepSeek R1 fully on CUDA with ik_llama.cpp @ dcdfad2.Layers are not evenly spread over 16 GPUs, and GPU utilisation is only at 5-10% on avg. <150w per GPU.
I'm not sure how useful this is, but ran it over night. It had an error on
-b 4096 pp8192due to OOM but still feel it's useful!Feel free to create whichever interesting graphs you find from it, as there's a lot of data it's quite hard to isolate:
PP
TG shows no notable difference.
Beta Was this translation helpful? Give feedback.
All reactions