Note
Go to the end to download the full example code.
Steel method forward to guess inputs and dynamic shapes (with Tiny-LLM)¶
Inputs are always dynamic with LLMs that is why dynamic shapes
needs to be specified when a LLM is exported with torch.export.export().
Most of the examples on HuggingFace use method
transformers.GenerationMixin.generate() but we only want to
export the model and its method forward.
That example shows to guess the inputs of this method even though the model
is executed through meth generate.
We focus on the model arnir0/Tiny-LLM. To avoid downloading any weights, we write a function creating a random model based on the same architecture.
Steel the forward method¶
The first step is to guess the dummy inputs. Let’s use the true model for that. We use the dummy example from the model page.
import copy
import pprint
import torch
import transformers
from onnx_diagnostic import doc
from onnx_diagnostic.helpers import string_type
from onnx_diagnostic.helpers.torch_test_helper import steal_forward
from onnx_diagnostic.torch_models.llms import get_tiny_llm
MODEL_NAME = "arnir0/Tiny-LLM"
tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)
model = transformers.AutoModelForCausalLM.from_pretrained(MODEL_NAME)
We rewrite the forward method to print the cache dimension.
def _forward_(*args, _f=None, **kwargs):
assert _f is not None
if not hasattr(torch.compiler, "is_exporting") or not torch.compiler.is_exporting():
# torch.compiler.is_exporting requires torch>=2.7
print("<-", string_type((args, kwargs), with_shape=True, with_min_max=True))
res = _f(*args, **kwargs)
if not hasattr(torch.compiler, "is_exporting") or not torch.compiler.is_exporting():
print("->", string_type(res, with_shape=True, with_min_max=True))
return res
keep_model_forward = model.forward
model.forward = lambda *args, _f=keep_model_forward, **kwargs: _forward_(
*args, _f=_f, **kwargs
)
Let’s run the model.
prompt = "Continue: it rains..."
inputs = tokenizer.encode(prompt, return_tensors="pt")
outputs = model.generate(
inputs, max_length=50, temperature=1, top_k=50, top_p=0.95, do_sample=True
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("-- prompt", prompt)
print("-- answer", generated_text)
<- ((),dict(cache_position:T7s8[0,7:A3.5],past_key_values:DynamicCache(key_cache=#0[], value_cache=#0[]),input_ids:T7s1x8[1,29901:A6305.375],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x8x32000[-15.516718864440918,15.75765609741211:A-3.381915190983544],past_key_values:DynamicCache(key_cache=#1[T1s1x1x8x96[-5.490959167480469,6.226877689361572:A-0.11321351693110653]], value_cache=#1[T1s1x1x8x96[-0.6787744760513306,0.49568021297454834:A0.007227749521139988]]))
<- ((),dict(cache_position:T7s1[8,8:A8.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x8x96[-5.490959167480469,6.226877689361572:A-0.11321351693110653]], value_cache=#1[T1s1x1x8x96[-0.6787744760513306,0.49568021297454834:A0.007227749521139988]]),input_ids:T7s1x1[13,13:A13.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-10.432564735412598,8.368535995483398:A-4.234468644971028],past_key_values:DynamicCache(key_cache=#1[T1s1x1x9x96[-5.509540557861328,6.348220348358154:A-0.12195695057461206]], value_cache=#1[T1s1x1x9x96[-0.6787744760513306,0.7704185843467712:A0.009565710057611594]]))
<- ((),dict(cache_position:T7s1[9,9:A9.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x9x96[-5.509540557861328,6.348220348358154:A-0.12195695057461206]], value_cache=#1[T1s1x1x9x96[-0.6787744760513306,0.7704185843467712:A0.009565710057611594]]),input_ids:T7s1x1[3629,3629:A3629.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-11.746018409729004,9.776823043823242:A-3.1345987820751033],past_key_values:DynamicCache(key_cache=#1[T1s1x1x10x96[-5.661314964294434,6.348220348358154:A-0.14218220768319345]], value_cache=#1[T1s1x1x10x96[-0.6787744760513306,0.7704185843467712:A0.008856571036797807]]))
<- ((),dict(cache_position:T7s1[10,10:A10.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x10x96[-5.661314964294434,6.348220348358154:A-0.14218220768319345]], value_cache=#1[T1s1x1x10x96[-0.6787744760513306,0.7704185843467712:A0.008856571036797807]]),input_ids:T7s1x1[292,292:A292.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.511404037475586,7.261352062225342:A-6.205918122938834],past_key_values:DynamicCache(key_cache=#1[T1s1x1x11x96[-5.661314964294434,6.348220348358154:A-0.13274977233904242]], value_cache=#1[T1s1x1x11x96[-0.6787744760513306,0.7704185843467712:A0.007832846533608861]]))
<- ((),dict(cache_position:T7s1[11,11:A11.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x11x96[-5.661314964294434,6.348220348358154:A-0.13274977233904242]], value_cache=#1[T1s1x1x11x96[-0.6787744760513306,0.7704185843467712:A0.007832846533608861]]),input_ids:T7s1x1[263,263:A263.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.752910614013672,3.912550926208496:A-7.536014426668873],past_key_values:DynamicCache(key_cache=#1[T1s1x1x12x96[-5.661314964294434,6.563353061676025:A-0.14228988134659126]], value_cache=#1[T1s1x1x12x96[-0.6787744760513306,0.7704185843467712:A0.00856697271864151]]))
<- ((),dict(cache_position:T7s1[12,12:A12.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x12x96[-5.661314964294434,6.563353061676025:A-0.14228988134659126]], value_cache=#1[T1s1x1x12x96[-0.6787744760513306,0.7704185843467712:A0.00856697271864151]]),input_ids:T7s1x1[285,285:A285.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-11.924566268920898,11.615700721740723:A-2.1173608495148364],past_key_values:DynamicCache(key_cache=#1[T1s1x1x13x96[-5.661314964294434,6.563353061676025:A-0.13358001891053087]], value_cache=#1[T1s1x1x13x96[-0.6787744760513306,0.7704185843467712:A0.008027582399394398]]))
<- ((),dict(cache_position:T7s1[13,13:A13.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x13x96[-5.661314964294434,6.563353061676025:A-0.13358001891053087]], value_cache=#1[T1s1x1x13x96[-0.6787744760513306,0.7704185843467712:A0.008027582399394398]]),input_ids:T7s1x1[1461,1461:A1461.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-13.426239013671875,15.311779022216797:A-2.3412022968861272],past_key_values:DynamicCache(key_cache=#1[T1s1x1x14x96[-5.661314964294434,6.563353061676025:A-0.11921060331527163]], value_cache=#1[T1s1x1x14x96[-0.6787744760513306,0.7704185843467712:A0.007556585159049757]]))
<- ((),dict(cache_position:T7s1[14,14:A14.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x14x96[-5.661314964294434,6.563353061676025:A-0.11921060331527163]], value_cache=#1[T1s1x1x14x96[-0.6787744760513306,0.7704185843467712:A0.007556585159049757]]),input_ids:T7s1x1[2955,2955:A2955.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-18.4624080657959,1.7358986139297485:A-9.963898456455208],past_key_values:DynamicCache(key_cache=#1[T1s1x1x15x96[-5.661314964294434,6.563353061676025:A-0.10007630323614447]], value_cache=#1[T1s1x1x15x96[-0.6787744760513306,0.7704185843467712:A0.006168467664560417]]))
<- ((),dict(cache_position:T7s1[15,15:A15.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x15x96[-5.661314964294434,6.563353061676025:A-0.10007630323614447]], value_cache=#1[T1s1x1x15x96[-0.6787744760513306,0.7704185843467712:A0.006168467664560417]]),input_ids:T7s1x1[11203,11203:A11203.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.273356437683105,7.016880035400391:A-7.015815305834636],past_key_values:DynamicCache(key_cache=#1[T1s1x1x16x96[-5.661314964294434,6.563353061676025:A-0.09172446005770023]], value_cache=#1[T1s1x1x16x96[-0.6787744760513306,0.7704185843467712:A0.005419444093215968]]))
<- ((),dict(cache_position:T7s1[16,16:A16.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x16x96[-5.661314964294434,6.563353061676025:A-0.09172446005770023]], value_cache=#1[T1s1x1x16x96[-0.6787744760513306,0.7704185843467712:A0.005419444093215968]]),input_ids:T7s1x1[393,393:A393.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.096044540405273,7.137486457824707:A-6.792652277584653],past_key_values:DynamicCache(key_cache=#1[T1s1x1x17x96[-5.661314964294434,6.563353061676025:A-0.09235078689387449]], value_cache=#1[T1s1x1x17x96[-0.6787744760513306,0.7704185843467712:A0.0057846603363032045]]))
<- ((),dict(cache_position:T7s1[17,17:A17.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x17x96[-5.661314964294434,6.563353061676025:A-0.09235078689387449]], value_cache=#1[T1s1x1x17x96[-0.6787744760513306,0.7704185843467712:A0.0057846603363032045]]),input_ids:T7s1x1[596,596:A596.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.924299240112305,9.052239418029785:A-7.81019115424226],past_key_values:DynamicCache(key_cache=#1[T1s1x1x18x96[-6.096094131469727,6.563353061676025:A-0.09724009151459204]], value_cache=#1[T1s1x1x18x96[-0.6787744760513306,0.7704185843467712:A0.006668706628724303]]))
<- ((),dict(cache_position:T7s1[18,18:A18.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x18x96[-6.096094131469727,6.563353061676025:A-0.09724009151459204]], value_cache=#1[T1s1x1x18x96[-0.6787744760513306,0.7704185843467712:A0.006668706628724303]]),input_ids:T7s1x1[11203,11203:A11203.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.362037658691406,9.275921821594238:A-7.130877728558262],past_key_values:DynamicCache(key_cache=#1[T1s1x1x19x96[-6.096094131469727,6.563353061676025:A-0.09020908581665585]], value_cache=#1[T1s1x1x19x96[-0.6787744760513306,0.7704185843467712:A0.006011621570530878]]))
<- ((),dict(cache_position:T7s1[19,19:A19.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x19x96[-6.096094131469727,6.563353061676025:A-0.09020908581665585]], value_cache=#1[T1s1x1x19x96[-0.6787744760513306,0.7704185843467712:A0.006011621570530878]]),input_ids:T7s1x1[1838,1838:A1838.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.466829299926758,13.748861312866211:A-6.65819999847142],past_key_values:DynamicCache(key_cache=#1[T1s1x1x20x96[-6.096094131469727,6.563353061676025:A-0.09111632193641223]], value_cache=#1[T1s1x1x20x96[-0.6787744760513306,0.7704185843467712:A0.005211767913647236]]))
<- ((),dict(cache_position:T7s1[20,20:A20.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x20x96[-6.096094131469727,6.563353061676025:A-0.09111632193641223]], value_cache=#1[T1s1x1x20x96[-0.6787744760513306,0.7704185843467712:A0.005211767913647236]]),input_ids:T7s1x1[29915,29915:A29915.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-19.425800323486328,17.70779037475586:A-5.884761626657099],past_key_values:DynamicCache(key_cache=#1[T1s1x1x21x96[-6.096094131469727,6.563353061676025:A-0.09017276051443852]], value_cache=#1[T1s1x1x21x96[-1.1154754161834717,0.7704185843467712:A0.0036792044997459995]]))
<- ((),dict(cache_position:T7s1[21,21:A21.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x21x96[-6.096094131469727,6.563353061676025:A-0.09017276051443852]], value_cache=#1[T1s1x1x21x96[-1.1154754161834717,0.7704185843467712:A0.0036792044997459995]]),input_ids:T7s1x1[29873,29873:A29873.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-19.159324645996094,8.502263069152832:A-7.592391656874679],past_key_values:DynamicCache(key_cache=#1[T1s1x1x22x96[-6.096094131469727,7.059808254241943:A-0.08803925118610839]], value_cache=#1[T1s1x1x22x96[-1.1154754161834717,0.7704185843467712:A0.00463082179002909]]))
<- ((),dict(cache_position:T7s1[22,22:A22.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x22x96[-6.096094131469727,7.059808254241943:A-0.08803925118610839]], value_cache=#1[T1s1x1x22x96[-1.1154754161834717,0.7704185843467712:A0.00463082179002909]]),input_ids:T7s1x1[505,505:A505.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-20.082448959350586,7.865069389343262:A-9.396246138746385],past_key_values:DynamicCache(key_cache=#1[T1s1x1x23x96[-6.096094131469727,7.059808254241943:A-0.08610944683465092]], value_cache=#1[T1s1x1x23x96[-1.1154754161834717,0.7704185843467712:A0.0029931723622963623]]))
<- ((),dict(cache_position:T7s1[23,23:A23.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x23x96[-6.096094131469727,7.059808254241943:A-0.08610944683465092]], value_cache=#1[T1s1x1x23x96[-1.1154754161834717,0.7704185843467712:A0.0029931723622963623]]),input_ids:T7s1x1[3307,3307:A3307.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-18.8045654296875,4.941972732543945:A-8.805901501461397],past_key_values:DynamicCache(key_cache=#1[T1s1x1x24x96[-6.096094131469727,7.059808254241943:A-0.08611314944305883]], value_cache=#1[T1s1x1x24x96[-1.1154754161834717,0.7704185843467712:A0.0029145279634311314]]))
<- ((),dict(cache_position:T7s1[24,24:A24.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x24x96[-6.096094131469727,7.059808254241943:A-0.08611314944305883]], value_cache=#1[T1s1x1x24x96[-1.1154754161834717,0.7704185843467712:A0.0029145279634311314]]),input_ids:T7s1x1[304,304:A304.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-14.548809051513672,9.828777313232422:A-5.6861186703937125],past_key_values:DynamicCache(key_cache=#1[T1s1x1x25x96[-6.731869697570801,7.059808254241943:A-0.08558527592778167]], value_cache=#1[T1s1x1x25x96[-1.1154754161834717,0.7704185843467712:A0.0034777971942518584]]))
<- ((),dict(cache_position:T7s1[25,25:A25.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x25x96[-6.731869697570801,7.059808254241943:A-0.08558527592778167]], value_cache=#1[T1s1x1x25x96[-1.1154754161834717,0.7704185843467712:A0.0034777971942518584]]),input_ids:T7s1x1[367,367:A367.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.280086517333984,7.913537979125977:A-6.349042969060596],past_key_values:DynamicCache(key_cache=#1[T1s1x1x26x96[-6.731869697570801,7.059808254241943:A-0.08352598655596338]], value_cache=#1[T1s1x1x26x96[-1.1154754161834717,0.7704185843467712:A0.003502177921122077]]))
<- ((),dict(cache_position:T7s1[26,26:A26.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x26x96[-6.731869697570801,7.059808254241943:A-0.08352598655596338]], value_cache=#1[T1s1x1x26x96[-1.1154754161834717,0.7704185843467712:A0.003502177921122077]]),input_ids:T7s1x1[263,263:A263.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.888399124145508,6.6775221824646:A-6.921862348470372],past_key_values:DynamicCache(key_cache=#1[T1s1x1x27x96[-6.731869697570801,7.059808254241943:A-0.08490211841768698]], value_cache=#1[T1s1x1x27x96[-1.1154754161834717,0.7704185843467712:A0.003988851359376839]]))
<- ((),dict(cache_position:T7s1[27,27:A27.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x27x96[-6.731869697570801,7.059808254241943:A-0.08490211841768698]], value_cache=#1[T1s1x1x27x96[-1.1154754161834717,0.7704185843467712:A0.003988851359376839]]),input_ids:T7s1x1[1559,1559:A1559.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.459266662597656,9.523560523986816:A-4.978380742765265],past_key_values:DynamicCache(key_cache=#1[T1s1x1x28x96[-6.731869697570801,7.059808254241943:A-0.0796025110880475]], value_cache=#1[T1s1x1x28x96[-1.1154754161834717,0.7704185843467712:A0.0037989136225394845]]))
<- ((),dict(cache_position:T7s1[28,28:A28.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x28x96[-6.731869697570801,7.059808254241943:A-0.0796025110880475]], value_cache=#1[T1s1x1x28x96[-1.1154754161834717,0.7704185843467712:A0.0037989136225394845]]),input_ids:T7s1x1[10963,10963:A10963.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.676679611206055,8.091779708862305:A-7.17866987408814],past_key_values:DynamicCache(key_cache=#1[T1s1x1x29x96[-6.731869697570801,7.059808254241943:A-0.06835894549081346]], value_cache=#1[T1s1x1x29x96[-1.1154754161834717,0.7704185843467712:A0.00370104361993866]]))
<- ((),dict(cache_position:T7s1[29,29:A29.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x29x96[-6.731869697570801,7.059808254241943:A-0.06835894549081346]], value_cache=#1[T1s1x1x29x96[-1.1154754161834717,0.7704185843467712:A0.00370104361993866]]),input_ids:T7s1x1[322,322:A322.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-15.402507781982422,6.966320037841797:A-7.022146884031594],past_key_values:DynamicCache(key_cache=#1[T1s1x1x30x96[-6.731869697570801,7.059808254241943:A-0.06671656767183676]], value_cache=#1[T1s1x1x30x96[-1.1154754161834717,0.7704185843467712:A0.0032558120854597897]]))
<- ((),dict(cache_position:T7s1[30,30:A30.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x30x96[-6.731869697570801,7.059808254241943:A-0.06671656767183676]], value_cache=#1[T1s1x1x30x96[-1.1154754161834717,0.7704185843467712:A0.0032558120854597897]]),input_ids:T7s1x1[285,285:A285.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-9.864608764648438,13.337820053100586:A-1.9672385881985537],past_key_values:DynamicCache(key_cache=#1[T1s1x1x31x96[-6.731869697570801,7.059808254241943:A-0.06289364934831786]], value_cache=#1[T1s1x1x31x96[-1.1154754161834717,0.7704185843467712:A0.003200943907491056]]))
<- ((),dict(cache_position:T7s1[31,31:A31.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x31x96[-6.731869697570801,7.059808254241943:A-0.06289364934831786]], value_cache=#1[T1s1x1x31x96[-1.1154754161834717,0.7704185843467712:A0.003200943907491056]]),input_ids:T7s1x1[391,391:A391.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.450756072998047,8.793631553649902:A-6.8214377727035895],past_key_values:DynamicCache(key_cache=#1[T1s1x1x32x96[-6.731869697570801,7.059808254241943:A-0.05801696362395129]], value_cache=#1[T1s1x1x32x96[-1.1154754161834717,0.7704185843467712:A0.0030400393721900323]]))
<- ((),dict(cache_position:T7s1[32,32:A32.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x32x96[-6.731869697570801,7.059808254241943:A-0.05801696362395129]], value_cache=#1[T1s1x1x32x96[-1.1154754161834717,0.7704185843467712:A0.0030400393721900323]]),input_ids:T7s1x1[975,975:A975.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-17.312728881835938,7.4546613693237305:A-7.682176516395761],past_key_values:DynamicCache(key_cache=#1[T1s1x1x33x96[-6.731869697570801,7.059808254241943:A-0.056958449662378174]], value_cache=#1[T1s1x1x33x96[-1.1154754161834717,0.7704185843467712:A0.0026938000898112614]]))
<- ((),dict(cache_position:T7s1[33,33:A33.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x33x96[-6.731869697570801,7.059808254241943:A-0.056958449662378174]], value_cache=#1[T1s1x1x33x96[-1.1154754161834717,0.7704185843467712:A0.0026938000898112614]]),input_ids:T7s1x1[278,278:A278.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-20.696626663208008,4.655063629150391:A-8.584791135491571],past_key_values:DynamicCache(key_cache=#1[T1s1x1x34x96[-6.731869697570801,7.059808254241943:A-0.05489631548110029]], value_cache=#1[T1s1x1x34x96[-1.1154754161834717,0.7704185843467712:A0.0030164116195357886]]))
<- ((),dict(cache_position:T7s1[34,34:A34.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x34x96[-6.731869697570801,7.059808254241943:A-0.05489631548110029]], value_cache=#1[T1s1x1x34x96[-1.1154754161834717,0.7704185843467712:A0.0030164116195357886]]),input_ids:T7s1x1[270,270:A270.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-9.699368476867676,14.263162612915039:A-1.4255934811742046],past_key_values:DynamicCache(key_cache=#1[T1s1x1x35x96[-6.731869697570801,7.059808254241943:A-0.054211245201232876]], value_cache=#1[T1s1x1x35x96[-1.1154754161834717,0.7704185843467712:A0.0033054195044704623]]))
<- ((),dict(cache_position:T7s1[35,35:A35.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x35x96[-6.731869697570801,7.059808254241943:A-0.054211245201232876]], value_cache=#1[T1s1x1x35x96[-1.1154754161834717,0.7704185843467712:A0.0033054195044704623]]),input_ids:T7s1x1[2335,2335:A2335.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.812931060791016,8.581151962280273:A-7.940854558266932],past_key_values:DynamicCache(key_cache=#1[T1s1x1x36x96[-6.731869697570801,7.059808254241943:A-0.049280428790148614]], value_cache=#1[T1s1x1x36x96[-1.1154754161834717,0.7704185843467712:A0.0028291172449986268]]))
<- ((),dict(cache_position:T7s1[36,36:A36.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x36x96[-6.731869697570801,7.059808254241943:A-0.049280428790148614]], value_cache=#1[T1s1x1x36x96[-1.1154754161834717,0.7704185843467712:A0.0028291172449986268]]),input_ids:T7s1x1[29889,29889:A29889.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.27593994140625,6.165885925292969:A-8.629735226717777],past_key_values:DynamicCache(key_cache=#1[T1s1x1x37x96[-6.731869697570801,7.074831485748291:A-0.049873085110546625]], value_cache=#1[T1s1x1x37x96[-1.1154754161834717,0.7704185843467712:A0.0030653254025842748]]))
<- ((),dict(cache_position:T7s1[37,37:A37.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x37x96[-6.731869697570801,7.074831485748291:A-0.049873085110546625]], value_cache=#1[T1s1x1x37x96[-1.1154754161834717,0.7704185843467712:A0.0030653254025842748]]),input_ids:T7s1x1[1670,1670:A1670.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-17.648948669433594,12.219874382019043:A-7.965085190537386],past_key_values:DynamicCache(key_cache=#1[T1s1x1x38x96[-6.731869697570801,7.074831485748291:A-0.049779349440812665]], value_cache=#1[T1s1x1x38x96[-1.1154754161834717,0.7704185843467712:A0.0018174540722922853]]))
<- ((),dict(cache_position:T7s1[38,38:A38.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x38x96[-6.731869697570801,7.074831485748291:A-0.049779349440812665]], value_cache=#1[T1s1x1x38x96[-1.1154754161834717,0.7704185843467712:A0.0018174540722922853]]),input_ids:T7s1x1[29915,29915:A29915.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-9.271356582641602,18.809484481811523:A-1.632277163193561],past_key_values:DynamicCache(key_cache=#1[T1s1x1x39x96[-6.731869697570801,7.074831485748291:A-0.048333697192241294]], value_cache=#1[T1s1x1x39x96[-1.1154754161834717,0.7704185843467712:A0.0010792613068930285]]))
<- ((),dict(cache_position:T7s1[39,39:A39.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x39x96[-6.731869697570801,7.074831485748291:A-0.048333697192241294]], value_cache=#1[T1s1x1x39x96[-1.1154754161834717,0.7704185843467712:A0.0010792613068930285]]),input_ids:T7s1x1[29879,29879:A29879.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-18.44225311279297,9.138684272766113:A-7.907652437511832],past_key_values:DynamicCache(key_cache=#1[T1s1x1x40x96[-6.731869697570801,7.404118537902832:A-0.04526488525355792]], value_cache=#1[T1s1x1x40x96[-1.1154754161834717,0.7704185843467712:A0.0012346806605781542]]))
<- ((),dict(cache_position:T7s1[40,40:A40.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x40x96[-6.731869697570801,7.404118537902832:A-0.04526488525355792]], value_cache=#1[T1s1x1x40x96[-1.1154754161834717,0.7704185843467712:A0.0012346806605781542]]),input_ids:T7s1x1[20947,20947:A20947.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-17.721210479736328,12.732599258422852:A-7.473342892593704],past_key_values:DynamicCache(key_cache=#1[T1s1x1x41x96[-6.731869697570801,7.404118537902832:A-0.04499705866384637]], value_cache=#1[T1s1x1x41x96[-1.1154754161834717,0.7704185843467712:A0.00034546581892437106]]))
<- ((),dict(cache_position:T7s1[41,41:A41.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x41x96[-6.731869697570801,7.404118537902832:A-0.04499705866384637]], value_cache=#1[T1s1x1x41x96[-1.1154754161834717,0.7704185843467712:A0.00034546581892437106]]),input_ids:T7s1x1[310,310:A310.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.469314575195312,6.776095867156982:A-5.998082326567965],past_key_values:DynamicCache(key_cache=#1[T1s1x1x42x96[-6.731869697570801,7.404118537902832:A-0.04336149620786951]], value_cache=#1[T1s1x1x42x96[-1.1154754161834717,0.7704185843467712:A0.0005079157287964913]]))
<- ((),dict(cache_position:T7s1[42,42:A42.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x42x96[-6.731869697570801,7.404118537902832:A-0.04336149620786951]], value_cache=#1[T1s1x1x42x96[-1.1154754161834717,0.7704185843467712:A0.0005079157287964913]]),input_ids:T7s1x1[26361,26361:A26361.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-14.865403175354004,9.806641578674316:A-6.575045243780594],past_key_values:DynamicCache(key_cache=#1[T1s1x1x43x96[-6.731869697570801,7.404118537902832:A-0.040619885459466244]], value_cache=#1[T1s1x1x43x96[-1.1154754161834717,0.7704185843467712:A0.0006090524590747869]]))
<- ((),dict(cache_position:T7s1[43,43:A43.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x43x96[-6.731869697570801,7.404118537902832:A-0.040619885459466244]], value_cache=#1[T1s1x1x43x96[-1.1154754161834717,0.7704185843467712:A0.0006090524590747869]]),input_ids:T7s1x1[363,363:A363.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-16.53607940673828,7.131961822509766:A-7.198457001279108],past_key_values:DynamicCache(key_cache=#1[T1s1x1x44x96[-6.731869697570801,7.758429527282715:A-0.0374408199152392]], value_cache=#1[T1s1x1x44x96[-1.1154754161834717,0.7704185843467712:A0.00045465556354773787]]))
<- ((),dict(cache_position:T7s1[44,44:A44.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x44x96[-6.731869697570801,7.758429527282715:A-0.0374408199152392]], value_cache=#1[T1s1x1x44x96[-1.1154754161834717,0.7704185843467712:A0.00045465556354773787]]),input_ids:T7s1x1[1781,1781:A1781.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-21.408689498901367,6.3573994636535645:A-8.284759337852709],past_key_values:DynamicCache(key_cache=#1[T1s1x1x45x96[-6.731869697570801,7.758429527282715:A-0.03618033421686657]], value_cache=#1[T1s1x1x45x96[-1.1154754161834717,0.7704185843467712:A0.0009134926121482244]]))
<- ((),dict(cache_position:T7s1[45,45:A45.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x45x96[-6.731869697570801,7.758429527282715:A-0.03618033421686657]], value_cache=#1[T1s1x1x45x96[-1.1154754161834717,0.7704185843467712:A0.0009134926121482244]]),input_ids:T7s1x1[286,286:A286.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-9.552547454833984,14.240614891052246:A-0.7839780572797171],past_key_values:DynamicCache(key_cache=#1[T1s1x1x46x96[-6.731869697570801,7.758429527282715:A-0.036942886177106025]], value_cache=#1[T1s1x1x46x96[-1.1154754161834717,0.7704185843467712:A0.0008465652450777798]]))
<- ((),dict(cache_position:T7s1[46,46:A46.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x46x96[-6.731869697570801,7.758429527282715:A-0.036942886177106025]], value_cache=#1[T1s1x1x46x96[-1.1154754161834717,0.7704185843467712:A0.0008465652450777798]]),input_ids:T7s1x1[1078,1078:A1078.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-14.190780639648438,10.620223999023438:A-6.825644922839012],past_key_values:DynamicCache(key_cache=#1[T1s1x1x47x96[-6.731869697570801,7.758429527282715:A-0.03455539851342406]], value_cache=#1[T1s1x1x47x96[-1.1154754161834717,0.7704185843467712:A0.00083209936923569]]))
<- ((),dict(cache_position:T7s1[47,47:A47.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x47x96[-6.731869697570801,7.758429527282715:A-0.03455539851342406]], value_cache=#1[T1s1x1x47x96[-1.1154754161834717,0.7704185843467712:A0.00083209936923569]]),input_ids:T7s1x1[322,322:A322.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-14.501764297485352,7.863560199737549:A-5.647588662019116],past_key_values:DynamicCache(key_cache=#1[T1s1x1x48x96[-6.731869697570801,7.758429527282715:A-0.03462720129136364]], value_cache=#1[T1s1x1x48x96[-1.1154754161834717,0.7704185843467712:A0.0006135993320760412]]))
<- ((),dict(cache_position:T7s1[48,48:A48.0],past_key_values:DynamicCache(key_cache=#1[T1s1x1x48x96[-6.731869697570801,7.758429527282715:A-0.03462720129136364]], value_cache=#1[T1s1x1x48x96[-1.1154754161834717,0.7704185843467712:A0.0006135993320760412]]),input_ids:T7s1x1[282,282:A282.0],inputs_embeds:None,use_cache:bool=True,return_dict:bool=True))
-> CausalLMOutputWithPast(logits:T1s1x1x32000[-11.180828094482422,16.256996154785156:A-1.1203185487377922],past_key_values:DynamicCache(key_cache=#1[T1s1x1x49x96[-6.731869697570801,7.758429527282715:A-0.031056391225202557]], value_cache=#1[T1s1x1x49x96[-1.1154754161834717,0.7704185843467712:A0.00010521440497875435]]))
-- prompt Continue: it rains...
-- answer Continue: it rains...
Being a fractured dog that your dog doesn't have enough to be a carpet and fist over the ditch. There's plenty of dogs for good mates and padd
Let’s restore the forward as it was.
model.forward = keep_model_forward
Another syntax with onnx_diagnostic.helpers.torch_test_helper.steal_forward().
with steal_forward(model):
model.generate(inputs, max_length=50, temperature=1, top_k=50, top_p=0.95, do_sample=True)
---- stolen forward for class LlamaForCausalLM -- iteration 0
<- args=() --- kwargs=dict(cache_position:T7s8,past_key_values:DynamicCache(key_cache=#0[], value_cache=#0[]),input_ids:T7s1x8,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x8x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x8x96], value_cache=#1[T1s1x1x8x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 1
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x8x96], value_cache=#1[T1s1x1x8x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x9x96], value_cache=#1[T1s1x1x9x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 2
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x9x96], value_cache=#1[T1s1x1x9x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x10x96], value_cache=#1[T1s1x1x10x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 3
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x10x96], value_cache=#1[T1s1x1x10x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x11x96], value_cache=#1[T1s1x1x11x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 4
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x11x96], value_cache=#1[T1s1x1x11x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x12x96], value_cache=#1[T1s1x1x12x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 5
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x12x96], value_cache=#1[T1s1x1x12x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x13x96], value_cache=#1[T1s1x1x13x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 6
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x13x96], value_cache=#1[T1s1x1x13x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x14x96], value_cache=#1[T1s1x1x14x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 7
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x14x96], value_cache=#1[T1s1x1x14x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x15x96], value_cache=#1[T1s1x1x15x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 8
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x15x96], value_cache=#1[T1s1x1x15x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x16x96], value_cache=#1[T1s1x1x16x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 9
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x16x96], value_cache=#1[T1s1x1x16x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x17x96], value_cache=#1[T1s1x1x17x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 10
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x17x96], value_cache=#1[T1s1x1x17x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x18x96], value_cache=#1[T1s1x1x18x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 11
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x18x96], value_cache=#1[T1s1x1x18x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x19x96], value_cache=#1[T1s1x1x19x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 12
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x19x96], value_cache=#1[T1s1x1x19x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x20x96], value_cache=#1[T1s1x1x20x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 13
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x20x96], value_cache=#1[T1s1x1x20x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x21x96], value_cache=#1[T1s1x1x21x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 14
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x21x96], value_cache=#1[T1s1x1x21x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x22x96], value_cache=#1[T1s1x1x22x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 15
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x22x96], value_cache=#1[T1s1x1x22x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x23x96], value_cache=#1[T1s1x1x23x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 16
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x23x96], value_cache=#1[T1s1x1x23x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x24x96], value_cache=#1[T1s1x1x24x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 17
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x24x96], value_cache=#1[T1s1x1x24x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x25x96], value_cache=#1[T1s1x1x25x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 18
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x25x96], value_cache=#1[T1s1x1x25x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x26x96], value_cache=#1[T1s1x1x26x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 19
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x26x96], value_cache=#1[T1s1x1x26x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x27x96], value_cache=#1[T1s1x1x27x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 20
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x27x96], value_cache=#1[T1s1x1x27x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x28x96], value_cache=#1[T1s1x1x28x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 21
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x28x96], value_cache=#1[T1s1x1x28x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x29x96], value_cache=#1[T1s1x1x29x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 22
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x29x96], value_cache=#1[T1s1x1x29x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x30x96], value_cache=#1[T1s1x1x30x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 23
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x30x96], value_cache=#1[T1s1x1x30x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x31x96], value_cache=#1[T1s1x1x31x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 24
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x31x96], value_cache=#1[T1s1x1x31x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x32x96], value_cache=#1[T1s1x1x32x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 25
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x32x96], value_cache=#1[T1s1x1x32x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x33x96], value_cache=#1[T1s1x1x33x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 26
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x33x96], value_cache=#1[T1s1x1x33x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x34x96], value_cache=#1[T1s1x1x34x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 27
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x34x96], value_cache=#1[T1s1x1x34x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x35x96], value_cache=#1[T1s1x1x35x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 28
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x35x96], value_cache=#1[T1s1x1x35x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x36x96], value_cache=#1[T1s1x1x36x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 29
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x36x96], value_cache=#1[T1s1x1x36x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x37x96], value_cache=#1[T1s1x1x37x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 30
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x37x96], value_cache=#1[T1s1x1x37x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x38x96], value_cache=#1[T1s1x1x38x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 31
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x38x96], value_cache=#1[T1s1x1x38x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x39x96], value_cache=#1[T1s1x1x39x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 32
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x39x96], value_cache=#1[T1s1x1x39x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x40x96], value_cache=#1[T1s1x1x40x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 33
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x40x96], value_cache=#1[T1s1x1x40x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x41x96], value_cache=#1[T1s1x1x41x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 34
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x41x96], value_cache=#1[T1s1x1x41x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x42x96], value_cache=#1[T1s1x1x42x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 35
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x42x96], value_cache=#1[T1s1x1x42x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x43x96], value_cache=#1[T1s1x1x43x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 36
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x43x96], value_cache=#1[T1s1x1x43x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x44x96], value_cache=#1[T1s1x1x44x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 37
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x44x96], value_cache=#1[T1s1x1x44x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x45x96], value_cache=#1[T1s1x1x45x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 38
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x45x96], value_cache=#1[T1s1x1x45x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x46x96], value_cache=#1[T1s1x1x46x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 39
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x46x96], value_cache=#1[T1s1x1x46x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x47x96], value_cache=#1[T1s1x1x47x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 40
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x47x96], value_cache=#1[T1s1x1x47x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x48x96], value_cache=#1[T1s1x1x48x96]))
.
---- stolen forward for class LlamaForCausalLM -- iteration 41
<- args=() --- kwargs=dict(cache_position:T7s1,past_key_values:DynamicCache(key_cache=#1[T1s1x1x48x96], value_cache=#1[T1s1x1x48x96]),input_ids:T7s1x1,inputs_embeds:None,use_cache:bool,return_dict:bool)
--
-> CausalLMOutputWithPast(logits:T1s1x1x32000,past_key_values:DynamicCache(key_cache=#1[T1s1x1x49x96], value_cache=#1[T1s1x1x49x96]))
.
Untrained model¶
This part can skipped if you are only interested in exporting the original model. It is useful to create a unit test to ensure a specific architecture can be exported despite the many changes brought to torch or transformers.
Let’s create an untrained model using the config file provided
config.json
to create an untrained model:
onnx_diagnostic.torch_models.llms.get_tiny_llm().
Then let’s use it.
experiment = get_tiny_llm()
untrained_model, inputs, dynamic_shapes = (
experiment["model"],
experiment["inputs"],
experiment["dynamic_shapes"],
)
Before we run it, we make a copy of the inputs as the cache get modified by the execution. Then it is no longer valid associated with the previous input_ids and mask.
print("input type before", string_type(inputs, with_shape=True))
expected_output = untrained_model(**inputs)
print("input type after-", string_type(inputs, with_shape=True))
input type before dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x30x96], value_cache=#1[T1s2x1x30x96]))
input type after- dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x33x96], value_cache=#1[T1s2x1x33x96]))
The outputs
print("result type", string_type(expected_output, with_shape=True))
result type CausalLMOutputWithPast(logits:T1s2x3x32000,past_key_values:DynamicCache(key_cache=#1[T1s2x1x33x96], value_cache=#1[T1s2x1x33x96]))
It works.
ExportedProgram¶
try:
ep = torch.export.export(
untrained_model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
)
print("It worked:")
print(ep)
except Exception as e:
# To work, it needs at least PRs:
# * https://github.com/huggingface/transformers/pull/36311
# * https://github.com/huggingface/transformers/pull/36652
print("It failed:", e)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s41, s2]", attention_mask: "i64[s41, s2 + s67]", position_ids: "i64[s41, s2]", past_key_values_key_cache_0: "f32[s41, 1, s67, 96]", past_key_values_value_cache_0: "f32[s41, 1, s67, 96]"):
#
sym_size_int_22: "Sym(s41)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_23: "Sym(s2)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_24: "Sym(s67)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s41, s2, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:542 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s2 + s67)" = sym_size_int_24 + sym_size_int_23
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:541 in forward, code: cache_position = torch.arange(
arange: "i64[s2]" = torch.ops.aten.arange.start(sym_size_int_24, add, device = device(type='cpu'), pin_memory = False); sym_size_int_24 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:548 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s2, s2 + s67]" = torch.ops.aten.full.default([sym_size_int_23, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s2, s2 + s67]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s2 + s67]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s2, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s2, s2 + s67]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s2, s2 + s67]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_22, 1, -1, -1]); slice_2 = None
clone: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_4: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_3, 1); slice_3 = None
slice_5: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_4, 2); slice_4 = None
slice_6: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_5, 3, None, add); slice_5 = None
slice_7: "i64[s41, s2 + s67]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s41, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(slice_7, 1); slice_7 = None
unsqueeze_3: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_8: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
_assert_tensor_metadata_default = torch.ops.aten._assert_tensor_metadata.default(slice_8, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default = None
to: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.to.dtype_layout(slice_8, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_8 = None
add_2: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.add.Tensor(slice_6, to); slice_6 = to = None
eq_4: "b8[s41, 1, s2, s2 + s67]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_9: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_10: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_9, 1); slice_9 = None
slice_11: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_10, 2); slice_10 = None
slice_12: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_11, 3, None, add); slice_11 = None
masked_fill: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.masked_fill.Scalar(slice_12, eq_4, -3.4028234663852886e+38); slice_12 = eq_4 = None
slice_13: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_14: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_13, 1, 0, 9223372036854775807); slice_13 = None
slice_15: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_14, 2, 0, 9223372036854775807); slice_14 = None
copy_: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.copy_.default(slice_15, masked_fill); slice_15 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_22, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:125 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[0]
to_7: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_8 = torch.ops.aten._assert_tensor_metadata.default(embedding, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_8 = None
to_8: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_8, 2)
mean: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_8, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_9 = torch.ops.aten._assert_tensor_metadata.default(mul_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_9 = None
to_9: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_9); p_model_layers_0_input_layernorm_weight = to_9 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_q_proj_weight); p_model_layers_0_self_attn_q_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:254 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s41, s2, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_22, sym_size_int_23, -1, 96]); linear = None
transpose_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_k_proj_weight); p_model_layers_0_self_attn_k_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:255 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_1 = None
transpose_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_v_proj_weight); mul_3 = p_model_layers_0_self_attn_v_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:256 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_2 = None
transpose_3: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:259 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
unsqueeze_8: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_7, 1); to_7 = None
mul_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_19: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_20: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s41, 2, s2, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.cat.default([neg, slice_19], -1); neg = slice_19 = None
mul_5: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_21: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_22: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s41, 1, s2, 48]" = torch.ops.aten.neg.default(slice_22); slice_22 = None
cat_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.cat.default([neg_1, slice_21], -1); neg_1 = slice_21 = None
mul_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s41, 1, s2, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:264 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:277 in forward, code: attn_output, attn_weights = attention_interface(
slice_23: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_24: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_23, 1, 0, 9223372036854775807); slice_23 = None
unsqueeze_9: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_24, 2); slice_24 = None
slice_25: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_26: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_25, 4, 0, 9223372036854775807); slice_25 = None
expand_2: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_26, [sym_size_int_22, 1, 2, add, 96]); slice_26 = None
reshape_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_22, 2, add, 96]); expand_2 = None
slice_27: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_28: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_27, 1, 0, 9223372036854775807); slice_27 = None
unsqueeze_10: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_28, 2); slice_28 = None
slice_29: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_30: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_29, 4, 0, 9223372036854775807); slice_29 = None
expand_3: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_30, [sym_size_int_22, 1, 2, add, 96]); slice_30 = None
reshape_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_22, 2, add, 96]); expand_3 = None
slice_31: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone); clone = None
slice_32: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_31, 1); slice_31 = None
slice_33: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_32, 2); slice_32 = None
slice_34: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_33, 3, None, add); slice_33 = add = None
contiguous: "f32[s41, 2, s2, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s41, 2, s2, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_34, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_34 = None
transpose_4: "f32[s41, s2, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s41, s2, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:288 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s41, s2, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_22, sym_size_int_23, -1]); contiguous_3 = sym_size_int_22 = sym_size_int_23 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:331 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_8, linear_3); to_8 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_10 = torch.ops.aten._assert_tensor_metadata.default(add_7, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_10 = None
to_10: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_10, 2)
mean_1: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_10, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_11 = torch.ops.aten._assert_tensor_metadata.default(mul_8, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_11 = None
to_11: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_11); p_model_layers_0_post_attention_layernorm_weight = to_11 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:434 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s41, s2, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:174 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s41, s2, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:337 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_10, linear_6); to_10 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_12 = torch.ops.aten._assert_tensor_metadata.default(add_9, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_12 = None
to_12: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_12, 2)
mean_2: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_12, rsqrt_2); to_12 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_13 = torch.ops.aten._assert_tensor_metadata.default(mul_11, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_13 = None
to_13: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_13); p_model_norm_weight = to_13 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:825 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_35: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(mul_12); mul_12 = None
slice_36: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_35, 1, 0); slice_35 = None
slice_37: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_36, 2); slice_36 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s41, s2, 32000]" = torch.ops.aten.linear.default(slice_37, p_lm_head_weight); slice_37 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_22: "Sym(s41)", position_ids: "i64[s41, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:115 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_16: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_16, 2); slice_16 = None
_assert_tensor_metadata_default_1 = torch.ops.aten._assert_tensor_metadata.default(unsqueeze_5, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_1 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s41, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_22, -1, 1]); to_1 = sym_size_int_22 = None
_assert_tensor_metadata_default_2 = torch.ops.aten._assert_tensor_metadata.default(expand_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_2 = None
to_2: "f32[s41, 48, 1]" = torch.ops.aten.to.dtype_layout(expand_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu')); expand_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:116 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_17: "i64[s41, s2]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s41, 1, s2]" = torch.ops.aten.unsqueeze.default(slice_17, 1); slice_17 = None
slice_18: "i64[s41, 1, s2]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
_assert_tensor_metadata_default_3 = torch.ops.aten._assert_tensor_metadata.default(slice_18, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_3 = None
to_3: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(slice_18, torch.float32); slice_18 = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_autocast = torch.ops.higher_order.wrap_with_autocast('cpu', torch.bfloat16, False, False, submod_3, to_2, to_3); submod_3 = to_2 = to_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:122 in forward, code: cos = emb.cos() * self.attention_scaling
mul: "f32[s41, s2, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:123 in forward, code: sin = emb.sin() * self.attention_scaling
mul_1: "f32[s41, s2, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:125 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
_assert_tensor_metadata_default_6 = torch.ops.aten._assert_tensor_metadata.default(mul, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_6 = None
to_6: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
_assert_tensor_metadata_default_7 = torch.ops.aten._assert_tensor_metadata.default(mul_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_7 = None
to_7: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_6, to_7)
class submod_1(torch.nn.Module):
def forward(self, to_2: "f32[s41, 48, 1]", to_3: "f32[s41, 1, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:120 in forward, code: freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
_assert_tensor_metadata_default_4 = torch.ops.aten._assert_tensor_metadata.default(to_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_4 = None
to_4: "f32[s41, 48, 1]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
_assert_tensor_metadata_default_5 = torch.ops.aten._assert_tensor_metadata.default(to_3, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_5 = None
to_5: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(to_3, torch.float32); to_3 = None
matmul: "f32[s41, 48, s2]" = torch.ops.aten.matmul.default(to_4, to_5); to_4 = to_5 = None
transpose: "f32[s41, s2, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:121 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s41, s2, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:122 in forward, code: cos = emb.cos() * self.attention_scaling
cos: "f32[s41, s2, 96]" = torch.ops.aten.cos.default(cat)
mul: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:123 in forward, code: sin = emb.sin() * self.attention_scaling
sin: "f32[s41, s2, 96]" = torch.ops.aten.sin.default(cat); cat = None
mul_1: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
return (mul, mul_1)
Graph signature:
# inputs
p_model_embed_tokens_weight: PARAMETER target='model.embed_tokens.weight'
p_model_layers_0_self_attn_q_proj_weight: PARAMETER target='model.layers.0.self_attn.q_proj.weight'
p_model_layers_0_self_attn_k_proj_weight: PARAMETER target='model.layers.0.self_attn.k_proj.weight'
p_model_layers_0_self_attn_v_proj_weight: PARAMETER target='model.layers.0.self_attn.v_proj.weight'
p_model_layers_0_self_attn_o_proj_weight: PARAMETER target='model.layers.0.self_attn.o_proj.weight'
p_model_layers_0_mlp_gate_proj_weight: PARAMETER target='model.layers.0.mlp.gate_proj.weight'
p_model_layers_0_mlp_up_proj_weight: PARAMETER target='model.layers.0.mlp.up_proj.weight'
p_model_layers_0_mlp_down_proj_weight: PARAMETER target='model.layers.0.mlp.down_proj.weight'
p_model_layers_0_input_layernorm_weight: PARAMETER target='model.layers.0.input_layernorm.weight'
p_model_layers_0_post_attention_layernorm_weight: PARAMETER target='model.layers.0.post_attention_layernorm.weight'
p_model_norm_weight: PARAMETER target='model.norm.weight'
p_lm_head_weight: PARAMETER target='lm_head.weight'
b_model_rotary_emb_inv_freq: BUFFER target='model.rotary_emb.inv_freq' persistent=False
input_ids: USER_INPUT
attention_mask: USER_INPUT
position_ids: USER_INPUT
past_key_values_key_cache_0: USER_INPUT
past_key_values_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s41: VR[1, 1024], s2: VR[2, 4096], s2 + s67: VR[4, 8192], s67: VR[1, 4096]}
Back to the original model¶
Let’s use the same dummy inputs but we use the downloaded model.
Dummy inputs and dynamic shapes are created by function
onnx_diagnostic.torch_models.llms.get_tiny_llm().
data = get_tiny_llm()
inputs, dynamic_shapes = data["inputs"], data["dynamic_shapes"]
Let’s print the inputs.
print(string_type(inputs, with_shape=True))
dict(input_ids:T7s2x3,attention_mask:T7s2x33,position_ids:T7s2x3,past_key_values:DynamicCache(key_cache=#1[T1s2x1x30x96], value_cache=#1[T1s2x1x30x96]))
{'attention_mask': {0: Dim('batch', min=1, max=1024),
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>,
min=None,
max=None,
_factory=True)},
'input_ids': {0: Dim('batch', min=1, max=1024),
1: Dim('seq_length', min=1, max=4096)},
'past_key_values': [[{0: Dim('batch', min=1, max=1024),
2: Dim('cache_length', min=1, max=4096)}],
[{0: Dim('batch', min=1, max=1024),
2: Dim('cache_length', min=1, max=4096)}]],
'position_ids': {0: Dim('batch', min=1, max=1024),
1: _DimHint(type=<_DimHintType.DYNAMIC: 3>,
min=None,
max=None,
_factory=True)}}
And Let’s finally export.
try:
ep = torch.export.export(
model, (), kwargs=cloned_inputs, dynamic_shapes=dynamic_shapes, strict=False
)
print("It worked:")
print(ep)
except Exception as e:
# To work, it needs at least PRs:
# * https://github.com/huggingface/transformers/pull/36311
# * https://github.com/huggingface/transformers/pull/36652
print("It failed:", e)
It worked:
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, p_model_embed_tokens_weight: "f32[32000, 192]", p_model_layers_0_self_attn_q_proj_weight: "f32[192, 192]", p_model_layers_0_self_attn_k_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_v_proj_weight: "f32[96, 192]", p_model_layers_0_self_attn_o_proj_weight: "f32[192, 192]", p_model_layers_0_mlp_gate_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_up_proj_weight: "f32[1024, 192]", p_model_layers_0_mlp_down_proj_weight: "f32[192, 1024]", p_model_layers_0_input_layernorm_weight: "f32[192]", p_model_layers_0_post_attention_layernorm_weight: "f32[192]", p_model_norm_weight: "f32[192]", p_lm_head_weight: "f32[32000, 192]", b_model_rotary_emb_inv_freq: "f32[48]", input_ids: "i64[s41, s2]", attention_mask: "i64[s41, s2 + s67]", position_ids: "i64[s41, s2]", past_key_values_key_cache_0: "f32[s41, 1, s67, 96]", past_key_values_value_cache_0: "f32[s41, 1, s67, 96]"):
#
sym_size_int_22: "Sym(s41)" = torch.ops.aten.sym_size.int(input_ids, 0)
sym_size_int_23: "Sym(s2)" = torch.ops.aten.sym_size.int(input_ids, 1)
sym_size_int_24: "Sym(s67)" = torch.ops.aten.sym_size.int(past_key_values_key_cache_0, 2)
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/sparse.py:190 in forward, code: return F.embedding(
embedding: "f32[s41, s2, 192]" = torch.ops.aten.embedding.default(p_model_embed_tokens_weight, input_ids); p_model_embed_tokens_weight = input_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:542 in forward, code: past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
add: "Sym(s2 + s67)" = sym_size_int_24 + sym_size_int_23
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:541 in forward, code: cache_position = torch.arange(
arange: "i64[s2]" = torch.ops.aten.arange.start(sym_size_int_24, add, device = device(type='cpu'), pin_memory = False); sym_size_int_24 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:548 in forward, code: causal_mask = self._update_causal_mask(
full: "f32[s2, s2 + s67]" = torch.ops.aten.full.default([sym_size_int_23, add], -3.4028234663852886e+38, dtype = torch.float32, device = device(type='cpu'), pin_memory = False)
triu: "f32[s2, s2 + s67]" = torch.ops.aten.triu.default(full, 1); full = None
arange_1: "i64[s2 + s67]" = torch.ops.aten.arange.default(add, device = device(type='cpu'), pin_memory = False)
reshape: "i64[s2, 1]" = torch.ops.aten.reshape.default(arange, [-1, 1]); arange = None
gt: "b8[s2, s2 + s67]" = torch.ops.aten.gt.Tensor(arange_1, reshape); arange_1 = reshape = None
mul_: "f32[s2, s2 + s67]" = torch.ops.aten.mul_.Tensor(triu, gt); triu = gt = None
unsqueeze: "f32[1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(mul_, 0); mul_ = None
unsqueeze_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze, 1); unsqueeze = None
slice_1: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_1, 2, 0, 9223372036854775807); unsqueeze_1 = None
slice_2: "f32[1, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_1, 3, 0, 9223372036854775807); slice_1 = None
expand: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.expand.default(slice_2, [sym_size_int_22, 1, -1, -1]); slice_2 = None
clone: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.clone.default(expand); expand = None
slice_3: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_4: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_3, 1); slice_3 = None
slice_5: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_4, 2); slice_4 = None
slice_6: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_5, 3, None, add); slice_5 = None
slice_7: "i64[s41, s2 + s67]" = torch.ops.aten.slice.Tensor(attention_mask, 0, 0, 9223372036854775807); attention_mask = None
unsqueeze_2: "i64[s41, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(slice_7, 1); slice_7 = None
unsqueeze_3: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.unsqueeze.default(unsqueeze_2, 2); unsqueeze_2 = None
slice_8: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.slice.Tensor(unsqueeze_3, 3, 0, 9223372036854775807); unsqueeze_3 = None
_assert_tensor_metadata_default = torch.ops.aten._assert_tensor_metadata.default(slice_8, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default = None
to: "i64[s41, 1, 1, s2 + s67]" = torch.ops.aten.to.dtype_layout(slice_8, dtype = torch.int64, layout = torch.strided, device = device(type='cpu')); slice_8 = None
add_2: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.add.Tensor(slice_6, to); slice_6 = to = None
eq_4: "b8[s41, 1, s2, s2 + s67]" = torch.ops.aten.eq.Scalar(add_2, 0); add_2 = None
slice_9: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone)
slice_10: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_9, 1); slice_9 = None
slice_11: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_10, 2); slice_10 = None
slice_12: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_11, 3, None, add); slice_11 = None
masked_fill: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.masked_fill.Scalar(slice_12, eq_4, -3.4028234663852886e+38); slice_12 = eq_4 = None
slice_13: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
slice_14: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_13, 1, 0, 9223372036854775807); slice_13 = None
slice_15: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_14, 2, 0, 9223372036854775807); slice_14 = None
copy_: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.copy_.default(slice_15, masked_fill); slice_15 = masked_fill = copy_ = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_set_grad_enabled = torch.ops.higher_order.wrap_with_set_grad_enabled(False, submod_3, b_model_rotary_emb_inv_freq, sym_size_int_22, position_ids); submod_3 = b_model_rotary_emb_inv_freq = position_ids = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:125 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
to_6: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[0]
to_7: "f32[s41, s2, 96]" = wrap_with_set_grad_enabled[1]; wrap_with_set_grad_enabled = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_8 = torch.ops.aten._assert_tensor_metadata.default(embedding, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_8 = None
to_8: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(embedding, torch.float32); embedding = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_1: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_8, 2)
mean: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_1, [-1], True); pow_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_3: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean, 1e-05); mean = None
rsqrt: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_3); add_3 = None
mul_2: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_8, rsqrt); rsqrt = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_9 = torch.ops.aten._assert_tensor_metadata.default(mul_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_9 = None
to_9: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_2, torch.float32); mul_2 = None
mul_3: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_input_layernorm_weight, to_9); p_model_layers_0_input_layernorm_weight = to_9 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_q_proj_weight); p_model_layers_0_self_attn_q_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:254 in forward, code: query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view: "f32[s41, s2, 2, 96]" = torch.ops.aten.view.default(linear, [sym_size_int_22, sym_size_int_23, -1, 96]); linear = None
transpose_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.transpose.int(view, 1, 2); view = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_1: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_k_proj_weight); p_model_layers_0_self_attn_k_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:255 in forward, code: key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_1: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_1, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_1 = None
transpose_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_1, 1, 2); view_1 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_2: "f32[s41, s2, 96]" = torch.ops.aten.linear.default(mul_3, p_model_layers_0_self_attn_v_proj_weight); mul_3 = p_model_layers_0_self_attn_v_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:256 in forward, code: value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
view_2: "f32[s41, s2, 1, 96]" = torch.ops.aten.view.default(linear_2, [sym_size_int_22, sym_size_int_23, -1, 96]); linear_2 = None
transpose_3: "f32[s41, 1, s2, 96]" = torch.ops.aten.transpose.int(view_2, 1, 2); view_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:259 in forward, code: query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
unsqueeze_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_6, 1); to_6 = None
unsqueeze_8: "f32[s41, 1, s2, 96]" = torch.ops.aten.unsqueeze.default(to_7, 1); to_7 = None
mul_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_1, unsqueeze_7)
slice_19: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 0, 48)
slice_20: "f32[s41, 2, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_1, 3, 48, 9223372036854775807); transpose_1 = None
neg: "f32[s41, 2, s2, 48]" = torch.ops.aten.neg.default(slice_20); slice_20 = None
cat_1: "f32[s41, 2, s2, 96]" = torch.ops.aten.cat.default([neg, slice_19], -1); neg = slice_19 = None
mul_5: "f32[s41, 2, s2, 96]" = torch.ops.aten.mul.Tensor(cat_1, unsqueeze_8); cat_1 = None
add_4: "f32[s41, 2, s2, 96]" = torch.ops.aten.add.Tensor(mul_4, mul_5); mul_4 = mul_5 = None
mul_6: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(transpose_2, unsqueeze_7); unsqueeze_7 = None
slice_21: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 0, 48)
slice_22: "f32[s41, 1, s2, 48]" = torch.ops.aten.slice.Tensor(transpose_2, 3, 48, 9223372036854775807); transpose_2 = None
neg_1: "f32[s41, 1, s2, 48]" = torch.ops.aten.neg.default(slice_22); slice_22 = None
cat_2: "f32[s41, 1, s2, 96]" = torch.ops.aten.cat.default([neg_1, slice_21], -1); neg_1 = slice_21 = None
mul_7: "f32[s41, 1, s2, 96]" = torch.ops.aten.mul.Tensor(cat_2, unsqueeze_8); cat_2 = unsqueeze_8 = None
add_5: "f32[s41, 1, s2, 96]" = torch.ops.aten.add.Tensor(mul_6, mul_7); mul_6 = mul_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:264 in forward, code: key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
cat_3: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_key_cache_0, add_5], -2); past_key_values_key_cache_0 = add_5 = None
cat_4: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.cat.default([past_key_values_value_cache_0, transpose_3], -2); past_key_values_value_cache_0 = transpose_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:277 in forward, code: attn_output, attn_weights = attention_interface(
slice_23: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_3, 0, 0, 9223372036854775807)
slice_24: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_23, 1, 0, 9223372036854775807); slice_23 = None
unsqueeze_9: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_24, 2); slice_24 = None
slice_25: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_9, 3, 0, 9223372036854775807); unsqueeze_9 = None
slice_26: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_25, 4, 0, 9223372036854775807); slice_25 = None
expand_2: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_26, [sym_size_int_22, 1, 2, add, 96]); slice_26 = None
reshape_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_2, [sym_size_int_22, 2, add, 96]); expand_2 = None
slice_27: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(cat_4, 0, 0, 9223372036854775807)
slice_28: "f32[s41, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_27, 1, 0, 9223372036854775807); slice_27 = None
unsqueeze_10: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.unsqueeze.default(slice_28, 2); slice_28 = None
slice_29: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(unsqueeze_10, 3, 0, 9223372036854775807); unsqueeze_10 = None
slice_30: "f32[s41, 1, 1, s2 + s67, 96]" = torch.ops.aten.slice.Tensor(slice_29, 4, 0, 9223372036854775807); slice_29 = None
expand_3: "f32[s41, 1, 2, s2 + s67, 96]" = torch.ops.aten.expand.default(slice_30, [sym_size_int_22, 1, 2, add, 96]); slice_30 = None
reshape_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.reshape.default(expand_3, [sym_size_int_22, 2, add, 96]); expand_3 = None
slice_31: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(clone); clone = None
slice_32: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_31, 1); slice_31 = None
slice_33: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_32, 2); slice_32 = None
slice_34: "f32[s41, 1, s2, s2 + s67]" = torch.ops.aten.slice.Tensor(slice_33, 3, None, add); slice_33 = add = None
contiguous: "f32[s41, 2, s2, 96]" = torch.ops.aten.contiguous.default(add_4); add_4 = None
contiguous_1: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_1); reshape_1 = None
contiguous_2: "f32[s41, 2, s2 + s67, 96]" = torch.ops.aten.contiguous.default(reshape_2); reshape_2 = None
scaled_dot_product_attention: "f32[s41, 2, s2, 96]" = torch.ops.aten.scaled_dot_product_attention.default(contiguous, contiguous_1, contiguous_2, slice_34, scale = 0.10206207261596575); contiguous = contiguous_1 = contiguous_2 = slice_34 = None
transpose_4: "f32[s41, s2, 2, 96]" = torch.ops.aten.transpose.int(scaled_dot_product_attention, 1, 2); scaled_dot_product_attention = None
contiguous_3: "f32[s41, s2, 2, 96]" = torch.ops.aten.contiguous.default(transpose_4); transpose_4 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:288 in forward, code: attn_output = attn_output.reshape(*input_shape, -1).contiguous()
reshape_3: "f32[s41, s2, 192]" = torch.ops.aten.reshape.default(contiguous_3, [sym_size_int_22, sym_size_int_23, -1]); contiguous_3 = sym_size_int_22 = sym_size_int_23 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_3: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(reshape_3, p_model_layers_0_self_attn_o_proj_weight); reshape_3 = p_model_layers_0_self_attn_o_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:331 in forward, code: hidden_states = residual + hidden_states
add_7: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_8, linear_3); to_8 = linear_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_10 = torch.ops.aten._assert_tensor_metadata.default(add_7, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_10 = None
to_10: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_7, torch.float32); add_7 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_2: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_10, 2)
mean_1: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_2, [-1], True); pow_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_8: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_1, 1e-05); mean_1 = None
rsqrt_1: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_8); add_8 = None
mul_8: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_10, rsqrt_1); rsqrt_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_11 = torch.ops.aten._assert_tensor_metadata.default(mul_8, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_11 = None
to_11: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_8, torch.float32); mul_8 = None
mul_9: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_layers_0_post_attention_layernorm_weight, to_11); p_model_layers_0_post_attention_layernorm_weight = to_11 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_4: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_gate_proj_weight); p_model_layers_0_mlp_gate_proj_weight = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/activation.py:434 in forward, code: return F.silu(input, inplace=self.inplace)
silu: "f32[s41, s2, 1024]" = torch.ops.aten.silu.default(linear_4); linear_4 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_5: "f32[s41, s2, 1024]" = torch.ops.aten.linear.default(mul_9, p_model_layers_0_mlp_up_proj_weight); mul_9 = p_model_layers_0_mlp_up_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:174 in forward, code: down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
mul_10: "f32[s41, s2, 1024]" = torch.ops.aten.mul.Tensor(silu, linear_5); silu = linear_5 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_6: "f32[s41, s2, 192]" = torch.ops.aten.linear.default(mul_10, p_model_layers_0_mlp_down_proj_weight); mul_10 = p_model_layers_0_mlp_down_proj_weight = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:337 in forward, code: hidden_states = residual + hidden_states
add_9: "f32[s41, s2, 192]" = torch.ops.aten.add.Tensor(to_10, linear_6); to_10 = linear_6 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:82 in forward, code: hidden_states = hidden_states.to(torch.float32)
_assert_tensor_metadata_default_12 = torch.ops.aten._assert_tensor_metadata.default(add_9, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_12 = None
to_12: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(add_9, torch.float32); add_9 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:83 in forward, code: variance = hidden_states.pow(2).mean(-1, keepdim=True)
pow_3: "f32[s41, s2, 192]" = torch.ops.aten.pow.Tensor_Scalar(to_12, 2)
mean_2: "f32[s41, s2, 1]" = torch.ops.aten.mean.dim(pow_3, [-1], True); pow_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:84 in forward, code: hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
add_10: "f32[s41, s2, 1]" = torch.ops.aten.add.Tensor(mean_2, 1e-05); mean_2 = None
rsqrt_2: "f32[s41, s2, 1]" = torch.ops.aten.rsqrt.default(add_10); add_10 = None
mul_11: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(to_12, rsqrt_2); to_12 = rsqrt_2 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:85 in forward, code: return self.weight * hidden_states.to(input_dtype)
_assert_tensor_metadata_default_13 = torch.ops.aten._assert_tensor_metadata.default(mul_11, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_13 = None
to_13: "f32[s41, s2, 192]" = torch.ops.aten.to.dtype(mul_11, torch.float32); mul_11 = None
mul_12: "f32[s41, s2, 192]" = torch.ops.aten.mul.Tensor(p_model_norm_weight, to_13); p_model_norm_weight = to_13 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:825 in forward, code: logits = self.lm_head(hidden_states[:, slice_indices, :])
slice_35: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(mul_12); mul_12 = None
slice_36: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_35, 1, 0); slice_35 = None
slice_37: "f32[s41, s2, 192]" = torch.ops.aten.slice.Tensor(slice_36, 2); slice_36 = None
# File: /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:125 in forward, code: return F.linear(input, self.weight, self.bias)
linear_7: "f32[s41, s2, 32000]" = torch.ops.aten.linear.default(slice_37, p_lm_head_weight); slice_37 = p_lm_head_weight = None
return (linear_7, cat_3, cat_4)
class submod_1(torch.nn.Module):
def forward(self, b_model_rotary_emb_inv_freq: "f32[48]", sym_size_int_22: "Sym(s41)", position_ids: "i64[s41, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:115 in forward, code: inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
unsqueeze_4: "f32[1, 48]" = torch.ops.aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0); b_model_rotary_emb_inv_freq = None
slice_16: "f32[1, 48]" = torch.ops.aten.slice.Tensor(unsqueeze_4, 1, 0, 9223372036854775807); unsqueeze_4 = None
unsqueeze_5: "f32[1, 48, 1]" = torch.ops.aten.unsqueeze.default(slice_16, 2); slice_16 = None
_assert_tensor_metadata_default_1 = torch.ops.aten._assert_tensor_metadata.default(unsqueeze_5, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_1 = None
to_1: "f32[1, 48, 1]" = torch.ops.aten.to.dtype(unsqueeze_5, torch.float32); unsqueeze_5 = None
expand_1: "f32[s41, 48, 1]" = torch.ops.aten.expand.default(to_1, [sym_size_int_22, -1, 1]); to_1 = sym_size_int_22 = None
_assert_tensor_metadata_default_2 = torch.ops.aten._assert_tensor_metadata.default(expand_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_2 = None
to_2: "f32[s41, 48, 1]" = torch.ops.aten.to.dtype_layout(expand_1, dtype = torch.float32, layout = torch.strided, device = device(type='cpu')); expand_1 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:116 in forward, code: position_ids_expanded = position_ids[:, None, :].float()
slice_17: "i64[s41, s2]" = torch.ops.aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807); position_ids = None
unsqueeze_6: "i64[s41, 1, s2]" = torch.ops.aten.unsqueeze.default(slice_17, 1); slice_17 = None
slice_18: "i64[s41, 1, s2]" = torch.ops.aten.slice.Tensor(unsqueeze_6, 2, 0, 9223372036854775807); unsqueeze_6 = None
_assert_tensor_metadata_default_3 = torch.ops.aten._assert_tensor_metadata.default(slice_18, dtype = torch.int64, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_3 = None
to_3: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(slice_18, torch.float32); slice_18 = None
# No stacktrace found for following nodes
submod_3 = self.submod_1
wrap_with_autocast = torch.ops.higher_order.wrap_with_autocast('cpu', torch.bfloat16, False, False, submod_3, to_2, to_3); submod_3 = to_2 = to_3 = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:122 in forward, code: cos = emb.cos() * self.attention_scaling
mul: "f32[s41, s2, 96]" = wrap_with_autocast[0]
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:123 in forward, code: sin = emb.sin() * self.attention_scaling
mul_1: "f32[s41, s2, 96]" = wrap_with_autocast[1]; wrap_with_autocast = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:125 in forward, code: return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
_assert_tensor_metadata_default_6 = torch.ops.aten._assert_tensor_metadata.default(mul, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_6 = None
to_6: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul, torch.float32); mul = None
_assert_tensor_metadata_default_7 = torch.ops.aten._assert_tensor_metadata.default(mul_1, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_7 = None
to_7: "f32[s41, s2, 96]" = torch.ops.aten.to.dtype(mul_1, torch.float32); mul_1 = None
return (to_6, to_7)
class submod_1(torch.nn.Module):
def forward(self, to_2: "f32[s41, 48, 1]", to_3: "f32[s41, 1, s2]"):
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:120 in forward, code: freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
_assert_tensor_metadata_default_4 = torch.ops.aten._assert_tensor_metadata.default(to_2, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_4 = None
to_4: "f32[s41, 48, 1]" = torch.ops.aten.to.dtype(to_2, torch.float32); to_2 = None
_assert_tensor_metadata_default_5 = torch.ops.aten._assert_tensor_metadata.default(to_3, dtype = torch.float32, device = device(type='cpu'), layout = torch.strided); _assert_tensor_metadata_default_5 = None
to_5: "f32[s41, 1, s2]" = torch.ops.aten.to.dtype(to_3, torch.float32); to_3 = None
matmul: "f32[s41, 48, s2]" = torch.ops.aten.matmul.default(to_4, to_5); to_4 = to_5 = None
transpose: "f32[s41, s2, 48]" = torch.ops.aten.transpose.int(matmul, 1, 2); matmul = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:121 in forward, code: emb = torch.cat((freqs, freqs), dim=-1)
cat: "f32[s41, s2, 96]" = torch.ops.aten.cat.default([transpose, transpose], -1); transpose = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:122 in forward, code: cos = emb.cos() * self.attention_scaling
cos: "f32[s41, s2, 96]" = torch.ops.aten.cos.default(cat)
mul: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(cos, 1.0); cos = None
# File: /home/xadupre/github/transformers/src/transformers/models/llama/modeling_llama.py:123 in forward, code: sin = emb.sin() * self.attention_scaling
sin: "f32[s41, s2, 96]" = torch.ops.aten.sin.default(cat); cat = None
mul_1: "f32[s41, s2, 96]" = torch.ops.aten.mul.Tensor(sin, 1.0); sin = None
return (mul, mul_1)
Graph signature:
# inputs
p_model_embed_tokens_weight: PARAMETER target='model.embed_tokens.weight'
p_model_layers_0_self_attn_q_proj_weight: PARAMETER target='model.layers.0.self_attn.q_proj.weight'
p_model_layers_0_self_attn_k_proj_weight: PARAMETER target='model.layers.0.self_attn.k_proj.weight'
p_model_layers_0_self_attn_v_proj_weight: PARAMETER target='model.layers.0.self_attn.v_proj.weight'
p_model_layers_0_self_attn_o_proj_weight: PARAMETER target='model.layers.0.self_attn.o_proj.weight'
p_model_layers_0_mlp_gate_proj_weight: PARAMETER target='model.layers.0.mlp.gate_proj.weight'
p_model_layers_0_mlp_up_proj_weight: PARAMETER target='model.layers.0.mlp.up_proj.weight'
p_model_layers_0_mlp_down_proj_weight: PARAMETER target='model.layers.0.mlp.down_proj.weight'
p_model_layers_0_input_layernorm_weight: PARAMETER target='model.layers.0.input_layernorm.weight'
p_model_layers_0_post_attention_layernorm_weight: PARAMETER target='model.layers.0.post_attention_layernorm.weight'
p_model_norm_weight: PARAMETER target='model.norm.weight'
p_lm_head_weight: PARAMETER target='lm_head.weight'
b_model_rotary_emb_inv_freq: BUFFER target='model.rotary_emb.inv_freq' persistent=False
input_ids: USER_INPUT
attention_mask: USER_INPUT
position_ids: USER_INPUT
past_key_values_key_cache_0: USER_INPUT
past_key_values_value_cache_0: USER_INPUT
# outputs
linear_7: USER_OUTPUT
cat_3: USER_OUTPUT
cat_4: USER_OUTPUT
Range constraints: {s41: VR[1, 1024], s2: VR[2, 4096], s2 + s67: VR[4, 8192], s67: VR[1, 4096]}
If you have any error, then look at example Export Tiny-LLM with patches.
doc.plot_legend("Tiny-LLM\nforward inputs\nbehind generate", "torch.export.export", "tomato")
Total running time of the script: (0 minutes 1.569 seconds)
Related examples
