Spaces:

Tinkering
/

Pytorch-day-prez

Running

App Files Files Community

Pytorch-day-prez / index.html

Molbap HF Staff

Update index.html

78bf448 verified 20 days ago

raw

history blame

27.9 kB

	<!DOCTYPE html>
	<html lang="en">
	<head>
	<meta charset="utf-8" />
	<meta name="viewport" content="width=device-width, initial-scale=1.0" />
	<title>PyTorch × Transformers Journey</title>

	<!-- Google Fonts -->
	<link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;600;800&family=Fira+Code:wght@400;600&display=swap" rel="stylesheet" />

	<!-- Reveal.js core & dark theme base -->
	<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/reveal.js@5/dist/reset.css" />
	<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/reveal.js@5/dist/reveal.css" />
	<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/reveal.js@5/dist/theme/black.css" id="theme" />

	<!-- Highlight.js -->
	<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/styles/github-dark.min.css" />

	<!-- Animations -->
	<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/animate.css@4/animate.min.css" />

	<style>
	:root {
	--accent-primary: #ee4c2c; /* PyTorch orange‑red */
	--accent-secondary: #ffb347; /* lighter highlight */
	--bg-gradient-start: #1b1b1b;
	--bg-gradient-end: #242424;
	}
	html, body { font-family: 'Inter', sans-serif; }
	.reveal .slides {
	background: linear-gradient(135deg, var(--bg-gradient-start), var(--bg-gradient-end));
	}
	.reveal h1, .reveal h2, .reveal h3 { color: var(--accent-primary); font-weight: 800; letter-spacing: -0.5px; }
	.reveal pre code { font-family: 'Fira Code', monospace; font-size: 0.75em; }
	.reveal section img, .reveal section svg { border-radius: 1rem; box-shadow: 0 8px 22px rgba(0,0,0,0.4); }
	.fragment.highlight-current-blue.visible { color: var(--accent-secondary) !important; }
	/* slide-density patch */
	.reveal h1 { font-size: 2.6rem; line-height: 1.1; }
	.reveal h2 { font-size: 1.9rem; line-height: 1.15; }
	.reveal h3 { font-size: 1.4rem; line-height: 1.2; }
	.reveal p, .reveal li { font-size: 1.7rem; line-height: 1.35; }
	.reveal pre code { font-size: 0.67em; }
	@media (max-width: 1024px) { .reveal h1{font-size:2.2rem;} .reveal h2{font-size:1.6rem;} }
	.reveal table td, .reveal table th { font-size: 0.85rem; padding: 4px 8px; }
	body::after {
	content: "";
	position: fixed;
	bottom: 3.5em;
	left: 3.5em;
	width: 270px; /* desired size */
	height: 117px;
	background-image: url(assets/py2.png);
	background-size: contain;
	background-repeat: no-repeat;
	z-index: 9999;
	box-shadow: 5px 5px 10px #000;
	pointer-events: none;
	}





	</style>
	</head>
	<body>
	<div class="reveal">
	<div class="slides">
	<section>
	<img src="assets/screenpage2.png" alt="Full slide image"
	style="
	width:120%;
	height:110%;
	object-fit:cover;
	margin-left:-2.5%;
	margin-top:-2.5%;
	" /> <!-- 1 · Opening -->
	</section>
	<section data-auto-animate>
	<img src="assets/head_logo.svg"
	alt="Logo"
	style="width: 120px; margin-bottom: 1rem;"
	class="animate__animated animate__fadeInDown" />
	<h1 class="animate__animated animate__fadeInDown">PyTorch × Transformers Journey</h1>
	<h3 class="animate__animated animate__fadeInDown animate__delay-1s">Pythonicity, Autodiff & Modularity in Modern AI</h3>
	<p class="animate__animated animate__fadeInUp animate__delay-2s">Pablo Montalvo‑Leroux  ·  ML Engineer @ Hugging Face</p>
	</section>

	<section>
	<h2>2016‑2018: Backprop & Birth Pangs</h2>
	<p>The journey began with uncertainty: back in 2016, machine learning was far from standardized. Tools like Theano and CNTK were fading, and many of us—myself included—were jumping framework to framework. It was a time of raw experimentation.</p>
	<ul>
	<li>Frameworks were in flux; few stuck around.</li>
	<li>MLPs evolved to RNNs and LSTMs.</li>
	<li><strong>2017, Attention, then 2018: BERT</strong> arrives, blowing the roof off what's possible.</li>
	</ul>
	<p class="fragment">But reproducing results remained frustratingly difficult.</p>
	</section>

	<section>
	<h2>Transformers × PyTorch: Reproducibility</h2>
	<p>That all changed with <code>pytorch-pretrained-bert</code>, the predecessor to Transformers. Suddenly, the magic of BERT was available in an interface that made sense.</p>
	<ul>
	<li>No static graphs, just Python functions and PyTorch modules.</li>
	<li>Readable, hackable code meant results could be shared, reproduced, improved.</li>
	<li>This shifted the research community towards PyTorch.</li>
	</ul>
	</section>


	<!-- 3 · Static vs Dynamic Graphs -->
	<section>
	<h2>Static vs Dynamic Graphs</h2>
	<p>Static graphs require you to compile, wait, and cross fingers the bug reproduces.</p>
	<p>Dynamic graphs mean you can drop <code>pdb.set_trace()</code> anywhere and continue iterating.</p>
	<p>Nowadays <code>torch.compile</code> gives the best of both worlds: write dynamically, ship something ahead‑of‑time optimised.</p>
	</section>


	<!-- 4 · Dynamic Graphs Enabled Contribution -->
	<section>
	<h2>Dynamic Graphs Enabled Contribution</h2>
	<ul>
	<li>Developers debug at line‑rate — no cold‑start recompiles.</li>
	<li>Pull‑requests remained reproducible overnight, which accelerated trust.</li>
	<li>Static‑graph alternatives stalled and the community consolidated around PyTorch.</li>
	</ul>
	</section>

	<section>
	<h2>Clone the Paper Tonight → Tweak Tomorrow</h2>
	<p>PyTorch lowered the barrier to implementation. Transformers removed the rest.</p>
	<ul>
	<li>2018: debugging BERT fine-tunes meant live tensor prints, not codegen restarts.</li>
	<li>Community credibility grew because patches could be merged fast and verified easily.</li>
	<li>Experimentation became a matter of hours, not weeks.</li>
	</ul>
	</section>

	<!-- 6 · One Model · One File -->
	<section>
	<h2>“One Model · One File” — Why it Matters</h2>
	<pre><code class="language-python" data-trim data-noescape>
	# modeling_bert.py — single source of truth
	class BertConfig(PretrainedConfig):
	...

	class BertSelfAttention(nn.Module):
	...

	class BertLayer(nn.Module):
	...

	class BertModel(PreTrainedModel):
	def __init__(self, config):
	super().__init__(config)
	self.embeddings = BertEmbeddings(config)
	self.encoder = nn.ModuleList(
	[BertLayer(config) for _ in range(config.num_hidden_layers)]
	)
	self.init_weights()
	</code></pre>
	<ul>
	<li>All layers, forward pass, and <code>from_pretrained()</code> logic live together.</li>
	<li>No cross‑file inheritance maze — copy to Colab, hack, and run.</li>
	<li>Reviewers diff one file; merge time dropped from days to hours.</li>
	</ul>
	</section>

	<section>
	<h2>Beyond Transformers: Ecosystem Reuse</h2>
	<p>Other libraries depend on <code>transformers</code> as a model definition source. For example, <strong>TRL</strong> uses models from the Hub directly:</p>

	<pre><code class="language-python" data-trim data-noescape>
	from datasets import load_dataset
	from transformers import AutoModelForCausalLM, AutoTokenizer
	from trl import DPOConfig, DPOTrainer

	model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
	tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
	dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
	training_args = DPOConfig(output_dir="Qwen2.5-0.5B-DPO")
	trainer = DPOTrainer(
	model=model,
	args=training_args,
	train_dataset=dataset,
	processing_class=tokenizer
	)
	trainer.train()
	</code></pre>

	<p class="fragment">No hacks, no refactoring — just <code>from_pretrained()</code>. Thanks to PyTorch autodiff and robust model definitions.</p>
	</section>


	<!-- 8 · Paradigms come at a cost -->
	<section>
	<h2>Paradigms come at a cost</h2>
	<ul>
	<p>The library took off, scientific and engineering ML community benefitted from it</p>
	<p>Torch adoption grew at the same time!</p>
	<p>The Hugging Face Hub became the AI app reference,</p>
	<p>In transformers, <strong> Maintenance</strong> becomes an issue: we have a lot of repeated code on purpose!</p>
	<p class="fragment">...but python is never far :) </p>

	</ul>
	</section>

	<!-- 8 · Back to Python: Mary Shelley Mode -->
	<section>
	<h2>Back to Python: Modular “Mary Shelley” Mode</h2>
	<p>Compose new blocks via subclass & override.</p>
	<pre><code class="language-python" data-trim>
	class GlmMLP(Phi3MLP):
	pass

	class GlmAttention(LlamaAttention):
	def __init__(self, config, layer_idx=None):
	super().__init__(config, layer_idx)
	self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim,
	config.hidden_size, bias=False)

	def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
	# Slightly different RoPE
	…

	class GlmForCausalLM(LlamaForCausalLM):
	pass
	</code></pre>
	<p>AST expands → full modeling file, still hackable.</p>
	</section>

	<section>
	<h2>Back to Python: It's alive!</h2>
	<p>All the code becomes runnable and a self-contained model definition</p>
	<pre><code class="language-python" data-trim>

	class GlmMLP(nn.Module):
	def __init__(self, config):
	super().__init__()

	self.config = config
	self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
	self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
	self.activation_fn = ACT2FN[config.hidden_act]

	def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
	up_states = self.gate_up_proj(hidden_states)

	gate, up_states = up_states.chunk(2, dim=-1)
	up_states = up_states * self.activation_fn(gate)

	return self.down_proj(up_states)


	def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
	"""
	This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
	num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
	"""
	batch, num_key_value_heads, slen, head_dim = hidden_states.shape
	if n_rep == 1:
	return hidden_states
	hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
	return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


	def eager_attention_forward(
	module: nn.Module,
	query: torch.Tensor,
	key: torch.Tensor,
	value: torch.Tensor,
	attention_mask: Optional[torch.Tensor],
	scaling: float,
	dropout: float = 0.0,
	**kwargs,
	):
	key_states = repeat_kv(key, module.num_key_value_groups)
	value_states = repeat_kv(value, module.num_key_value_groups)

	attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
	if attention_mask is not None:
	causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
	attn_weights = attn_weights + causal_mask

	attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
	attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
	attn_output = torch.matmul(attn_weights, value_states)
	attn_output = attn_output.transpose(1, 2).contiguous()

	return attn_output, attn_weights


	def rotate_half(x):
	"""Rotates half the hidden dims of the input."""
	x1 = x[..., 0::2]
	x2 = x[..., 1::2]
	return torch.stack((-x2, x1), dim=-1).flatten(-2)


	def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
	"""Applies Rotary Position Embedding to the query and key tensors.

	Args:
	q (`torch.Tensor`): The query tensor.
	k (`torch.Tensor`): The key tensor.
	cos (`torch.Tensor`): The cosine part of the rotary embedding.
	sin (`torch.Tensor`): The sine part of the rotary embedding.
	position_ids (`torch.Tensor`, optional):
	Deprecated and unused.
	unsqueeze_dim (`int`, optional, defaults to 1):
	The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
	sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
	that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
	k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
	cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
	the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
	Returns:
	`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
	"""
	cos = cos.unsqueeze(unsqueeze_dim)
	sin = sin.unsqueeze(unsqueeze_dim)

	# Interleave them instead of usual shape
	cos = cos[..., : cos.shape[-1] // 2].repeat_interleave(2, dim=-1)
	sin = sin[..., : sin.shape[-1] // 2].repeat_interleave(2, dim=-1)

	# Keep half or full tensor for later concatenation
	rotary_dim = cos.shape[-1]
	q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
	k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

	# Apply rotary embeddings on the first half or full tensor
	q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
	k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)

	# Concatenate back to full shape
	q_embed = torch.cat([q_embed, q_pass], dim=-1)
	k_embed = torch.cat([k_embed, k_pass], dim=-1)
	return q_embed, k_embed


	class GlmAttention(nn.Module):
	"""Multi-headed attention from 'Attention Is All You Need' paper"""

	def __init__(self, config: GlmConfig, layer_idx: Optional[int] = None):
	super().__init__()
	self.config = config
	self.layer_idx = layer_idx
	self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
	self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
	self.scaling = self.head_dim**-0.5
	self.attention_dropout = config.attention_dropout
	self.is_causal = True

	self.q_proj = nn.Linear(
	config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
	)
	self.k_proj = nn.Linear(
	config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
	)
	self.v_proj = nn.Linear(
	config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
	)
	self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)

	def forward(
	self,
	hidden_states: torch.Tensor,
	position_embeddings: Tuple[torch.Tensor, torch.Tensor],
	attention_mask: Optional[torch.Tensor],
	past_key_value: Optional[Cache] = None,
	cache_position: Optional[torch.LongTensor] = None,
	**kwargs: Unpack[FlashAttentionKwargs],
	) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
	input_shape = hidden_states.shape[:-1]
	hidden_shape = (*input_shape, -1, self.head_dim)

	query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
	key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
	value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

	cos, sin = position_embeddings
	query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

	if past_key_value is not None:
	# sin and cos are specific to RoPE models; cache_position needed for the static cache
	cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
	key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

	attention_interface: Callable = eager_attention_forward

	if self.config._attn_implementation != "eager":
	if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
	logger.warning_once(
	"`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
	'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
	)
	else:
	attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

	attn_output, attn_weights = attention_interface(
	self,
	query_states,
	key_states,
	value_states,
	attention_mask,
	dropout=0.0 if not self.training else self.attention_dropout,
	scaling=self.scaling,
	**kwargs,
	)

	attn_output = attn_output.reshape(*input_shape, -1).contiguous()
	attn_output = self.o_proj(attn_output)
	return attn_output, attn_weights


	@use_kernel_forward_from_hub("RMSNorm")
	class GlmRMSNorm(nn.Module):
	def __init__(self, hidden_size, eps=1e-6):
	"""
	GlmRMSNorm is equivalent to T5LayerNorm
	"""
	super().__init__()
	self.weight = nn.Parameter(torch.ones(hidden_size))
	self.variance_epsilon = eps

	def forward(self, hidden_states):
	input_dtype = hidden_states.dtype
	hidden_states = hidden_states.to(torch.float32)
	variance = hidden_states.pow(2).mean(-1, keepdim=True)
	hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
	return self.weight * hidden_states.to(input_dtype)

	def extra_repr(self):
	return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


	class GlmRotaryEmbedding(nn.Module):
	def __init__(self, config: GlmConfig, device=None):
	super().__init__()
	# BC: "rope_type" was originally "type"
	if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
	self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
	else:
	self.rope_type = "default"
	self.max_seq_len_cached = config.max_position_embeddings
	self.original_max_seq_len = config.max_position_embeddings

	self.config = config
	self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

	inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
	self.register_buffer("inv_freq", inv_freq, persistent=False)
	self.original_inv_freq = self.inv_freq
	</code></pre>
	<p> We keep hackability while reconnecting with Python working paradigms.</p>
	</section>


	<!-- 9 · Logit Debugger -->
	<section>
	<h2>Logit Debugger: Trust but Verify</h2>
	<ul>
	<li>Hook every <code>nn.Module</code>; dump logits layer‑by‑layer</li>
	<li>Spot ε‑level drifts (LayerNorm, FP16 underflow…)</li>
	<li>JSON traces diffable in CI</li>
	<img data-src="assets/visual_debugger.png" alt="Visual debugger" />

	</ul>
	</section>

	<!-- 10 · DTensor & TP API -->
	<section>
	<h2>DTensor & Tensor‑Parallel API</h2>
	<p>Before, changing to Tensor Parallel meant changing the code.</p>

	<pre><code class="language-python" data-trim data-noescape>
	from transformers.modeling_utils import PreTrainedModel
	from megatron.model import ColumnParallelLinear, RowParallelLinear

	class MyTPModel(PreTrainedModel):
	def __init__(self, config):
	super().__init__(config)
	self.q_proj = ColumnParallelLinear(config.hidden_size, config.hidden_size)
	self.k_proj = ColumnParallelLinear(config.hidden_size, config.hidden_size)
	self.v_proj = ColumnParallelLinear(config.hidden_size, config.hidden_size)
	self.o_proj = RowParallelLinear(config.hidden_size, config.hidden_size)

	</code></pre>
	</section>

	<!-- 11 · Zero‑Config Parallelism -->
	<section>
	<h2>Zero‑Config Tensor Parallelism</h2>
	<p>The <code>tp_plan</code> JSON keeps model code pristine and declarative.</p>
	<pre><code class="language-json" data-trim data-noescape>{
	"layer.*.self_attn.q_proj": "colwise",
	"layer.*.self_attn.k_proj": "colwise",
	"layer.*.self_attn.v_proj": "colwise",
	"layer.*.self_attn.o_proj": "rowwise"
	}</code></pre>
	<p class="fragment">Translated to</p>

	<pre><code class="language-python" data-trim data-noescape>
	def translate_to_torch_parallel_style(style: str):
	if style == "colwise":
	return ColwiseParallel()
	elif style == "rowwise":
	return RowwiseParallel()
	# …
	</code></pre>
	<p class="fragment">One JSON → 100 B param model on 8 GPUs. Change the plan, not the code.</p>
	</section>

	<!-- 12 · Cache Allocator -->
	<section>
	<h2>Improvements, Load faster & stronger: Cache Allocator</h2>
	<p>0‑copy weight sharding, single cuda Malloc</p>
	<p>Faster model loads, even for a 50-shards 100B model (when we were sprinting Llama4!)</p>
	<img data-src="assets/fastload.png" alt="SurprisedLewis" />
	</section>

	<!-- 15 · Why Python wins -->
	<section>
	<h2>Why Python Wins</h2>
	<ul>
	<li>Low entry barrier (although hard to master)</li>
	<li>High‑level semantics express low‑level intent</li>
	<li>Seamless C++/Rust extension points</li>
	</ul>
	</section>

	<!-- 16 · Where Python can bite -->
	<section>
	<h2>Where Python can bite 🐍</h2>
	<ul>
	<li>Interpreter overhead on microkernels (token‑by‑token decode)</li>
	<li>GIL can throttle async host‑side work</li>
	<li>Easy to under‑optimise code fresh out of the lab</li>
	</ul>
	<p class="fragment">All of these can be mitigated: Triton, compiled custom ops, compile‑time fallback, <strong>custom kernels</strong></p>
	</section>


	<!-- 17 · Kernel Hub -->
	<section>
	<h2>Kernel Hub: Optimised Ops from the Community</h2>
	<p>Kernel Hub lets any Python program <em>download and hot‑load</em> compiled CUDA/C++ kernels directly from the Hugging Face Hub at runtime.</p>
	<ul>
	<li><strong>Portable</strong> – kernels work from arbitrary paths outside <code>PYTHONPATH</code>.</li>
	<li><strong>Unique</strong> – load multiple versions of the same op side‑by‑side in one process.</li>
	<li><strong>Compatible</strong> – every kernel targets all recent PyTorch wheels (CUDA, ROCm, CPU) and C‑library ABIs.</li>
	</ul>
	<pre><code class="language-python" data-trim data-noescape>
	import torch
	from kernels import get_kernel

	# Download optimised kernels from the Hugging Face Hub
	activation = get_kernel("kernels-community/activation")

	x = torch.randn(10, 10, dtype=torch.float16, device="cuda")
	y = torch.empty_like(x)
	activation.gelu_fast(y, x)
	print(y)
	</code></pre>
	<p class="fragment">Same Transformer code — now with a <strong>3× faster</strong> GELU on A100s.</p>
	</section>


	<!-- 18 · API design lessons -->
	<section>
	<h2>API Design Lessons</h2>
	<ul>
	<li>Make easy things obvious, hard things possible</li>
	<li>Paper‑to‑repo diff should be minimal</li>
	<li>Research repo-to-stable architecture should be as fast as possible</li>

	<li>Hide sharding, expose intent</li>
	</ul>
	<p>We tune radios without building RF amplifiers — ML should feel the same.</p>
	<p class="fragment">..while enabling people who build the amplifiers.</p>

	</section>


	<!-- 14 · Rise of Multimodality -->
	<section>
	<h2>Rise of Multimodality</h2>
	<pre><code class="language-python" data-trim data-noescape>
	processor = AutoProcessor.from_pretrained("Qwen/Qwen3-8B")
	model = AutoModelForConditionalGeneration.from_pretrained("Qwen/Qwen3-8B")
	</code></pre>
	<p>Same API across text · vision · audio</p>
	<p>More and more models, with specific processing - need to uniformize</p>

	</section>

	<section>
	<h2>Rise of Multimodality: torch-powered processing</h2>
	<p>Torch and torchvision ops have replaced np + PIL defaults in transformers</p>

	<img data-src="assets/normalize_time_torch.webp" width="80%" height="600" alt="Fast load" />

	</section>
	<!-- 19 · Model Growth by Modality -->
	<section>
	<h2>Model Growth by Modality</h2>
	<iframe src="assets/model_growth.html" width="80%" height="600" style="border:none;"></iframe>
	</section>

	<!-- 20 · Takeaways -->
	<section>
	<h2>Takeaways & The Future</h2>
	<ul>
	<li style="display: flex; align-items: center; gap: 1rem;">
	<img src="assets/torchlogo.png" alt="PyTorch" style="height: 2rem;" />
	PyTorch & <code>transformers</code> grow symbiotically
	<img src="assets/head_logo.svg" alt="Transformers" style="height: 2rem;" />
	</li>
	<li>Pythonicity × pragmatism drive adoption</li>
	<li>Open‑source models are shipping faster & bigger than ever</li>
	<li class="fragment"> Let's go!</li>

	</ul>
	<p>
	<a href="https://huggingface.co/transformers/contribute" target="_blank">
	hf.co/transformers/contribute
	</a>
	</p>
	</section>


	</div>
	</div>

	<!-- Reveal.js core -->
	<script src="https://cdn.jsdelivr.net/npm/reveal.js@5/dist/reveal.js"></script>
	<script src="https://cdn.jsdelivr.net/npm/reveal.js@5/plugin/highlight/highlight.js"></script>
	<script src="https://cdn.jsdelivr.net/npm/reveal.js@5/plugin/notes/notes.js"></script>
	<!-- Plotly for interactive charts -->
	<script src="https://cdn.plot.ly/plotly-2.31.1.min.js"></script>
	<script>
	Reveal.initialize({
	hash: true,
	slideNumber: true,
	transition: 'slide',
	backgroundTransition: 'convex',
	plugins: [ RevealHighlight, RevealNotes ]
	});
	</script>
	</body>
	</html>