#!/usr/bin/env python3 # -*- coding: utf-8 -*- """AMD AAI2026 开发者版 12 页 PPT 受众：AI 工程师 / ML 平台开发者 / 推理服务开发者 风格：代码 / CLI / 终端友好，黑底代码块 + 浅底说明 输出：/var/www/html/share/AMD-AAI2026-演示-开发者版.pptx """ import os from pptx import Presentation from pptx.util import Inches, Pt from pptx.dml.color import RGBColor from pptx.enum.shapes import MSO_SHAPE from pptx.enum.text import PP_ALIGN, MSO_ANCHOR from pptx.oxml.ns import qn from lxml import etree # ---- 配色（开发者版：终端风 + AMD 红点缀）---- AMD_RED = RGBColor(0xED, 0x1C, 0x24) NV_GREEN = RGBColor(0x76, 0xB9, 0x00) DARK = RGBColor(0x1A, 0x1F, 0x2A) DARKER = RGBColor(0x0E, 0x11, 0x18) TERMINAL = RGBColor(0x10, 0x14, 0x1C) TERM_FG = RGBColor(0xE6, 0xE9, 0xEF) TERM_CMT = RGBColor(0x8A, 0x9B, 0xA8) TERM_KEY = RGBColor(0xFF, 0xD2, 0x8A) TERM_STR = RGBColor(0xA9, 0xDC, 0x9E) TERM_OK = RGBColor(0x76, 0xB9, 0x00) TERM_RED = RGBColor(0xFF, 0x8A, 0x8A) GREY = RGBColor(0x5A, 0x64, 0x78) LIGHT = RGBColor(0xE7, 0xEA, 0xF0) PANEL = RGBColor(0xF7, 0xF8, 0xFA) WHITE = RGBColor(0xFF, 0xFF, 0xFF) WARN = RGBColor(0xD4, 0x4A, 0x3A) CN_FONT = "Noto Sans CJK SC" OUT_PPTX = "/var/www/html/share/AMD-AAI2026-演示-开发者版.pptx" SLIDE_W = Inches(13.333) SLIDE_H = Inches(7.5) def add_rect(slide, x, y, w, h, fill, line=None, line_w=None): shp = slide.shapes.add_shape(MSO_SHAPE.RECTANGLE, x, y, w, h) shp.fill.solid(); shp.fill.fore_color.rgb = fill if line is None: shp.line.fill.background() else: shp.line.color.rgb = line if line_w is not None: shp.line.width = line_w shp.shadow.inherit = False return shp def add_text(slide, x, y, w, h, text, *, size=18, bold=False, color=DARK, align=PP_ALIGN.LEFT, anchor=MSO_ANCHOR.TOP, font=CN_FONT): tb = slide.shapes.add_textbox(x, y, w, h) tf = tb.text_frame tf.margin_left = Inches(0.08); tf.margin_right = Inches(0.08) tf.margin_top = Inches(0.04); tf.margin_bottom = Inches(0.04) tf.word_wrap = True tf.vertical_anchor = anchor lines = text.split("\n") if isinstance(text, str) else [str(text)] for i, ln in enumerate(lines): p = tf.paragraphs[0] if i == 0 else tf.add_paragraph() p.alignment = align run = p.add_run() run.text = ln run.font.size = Pt(size) run.font.bold = bold run.font.name = font rPr = run._r.get_or_add_rPr() ea = rPr.find(qn('a:ea')) if ea is None: ea = etree.SubElement(rPr, qn('a:ea')) ea.set('typeface', CN_FONT) run.font.color.rgb = color return tb def add_title_bar(slide, title, subtitle=None): """开发者版标题条：白底 + 黑色标题 + 红色左侧条""" add_rect(slide, 0, 0, SLIDE_W, Inches(0.9), WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.75)) add_rect(slide, 0, 0, Inches(0.15), Inches(0.9), AMD_RED) add_text(slide, Inches(0.4), Inches(0.1), SLIDE_W - Inches(1), Inches(0.55), title, size=24, bold=True, color=DARK, anchor=MSO_ANCHOR.MIDDLE) if subtitle: add_text(slide, Inches(0.4), Inches(0.5), SLIDE_W - Inches(1), Inches(0.38), subtitle, size=11, color=GREY, anchor=MSO_ANCHOR.TOP) def add_footer(slide, page_num, total=12): add_rect(slide, 0, SLIDE_H - Inches(0.35), SLIDE_W, Inches(0.35), RGBColor(0xF2, 0xF3, 0xF6)) add_text(slide, Inches(0.4), SLIDE_H - Inches(0.32), Inches(11), Inches(0.3), "适用 ROCm 7.0+ · MI455X · EPYC Venice · 开发者文档持续更新", size=9, color=GREY, anchor=MSO_ANCHOR.MIDDLE) add_text(slide, SLIDE_W - Inches(1.0), SLIDE_H - Inches(0.32), Inches(0.7), Inches(0.3), f"{page_num} / {total}", size=10, color=GREY, align=PP_ALIGN.RIGHT, anchor=MSO_ANCHOR.MIDDLE) def code_block(slide, x, y, w, h, lines, *, title=None, font_size=12): """终端风格的代码块：黑底 + 颜色高亮 + 等宽字体 lines: list of (text, color_tag) color_tag ∈ 'c'/'k'/'s'/'o'/'r'/None """ if title: # 顶部标签条 add_rect(slide, x, y, w, Inches(0.3), RGBColor(0x1E, 0x24, 0x33)) add_text(slide, x + Inches(0.1), y, w - Inches(0.2), Inches(0.3), title, size=10, color=TERM_FG, anchor=MSO_ANCHOR.MIDDLE) y = y + Inches(0.3) h = h - Inches(0.3) add_rect(slide, x, y, w, h, TERMINAL, line=RGBColor(0x2A, 0x32, 0x44), line_w=Pt(0.5)) tb = slide.shapes.add_textbox(x + Inches(0.1), y + Inches(0.05), w - Inches(0.2), h - Inches(0.1)) tf = tb.text_frame tf.margin_left = Inches(0); tf.margin_right = Inches(0) tf.margin_top = Inches(0); tf.margin_bottom = Inches(0) tf.word_wrap = True for i, item in enumerate(lines): if isinstance(item, tuple): text, tag = item else: text, tag = item, None p = tf.paragraphs[0] if i == 0 else tf.add_paragraph() p.alignment = PP_ALIGN.LEFT run = p.add_run() run.text = text run.font.size = Pt(font_size) run.font.name = "Courier New" rPr = run._r.get_or_add_rPr() ea = rPr.find(qn('a:ea')) if ea is None: ea = etree.SubElement(rPr, qn('a:ea')) ea.set('typeface', CN_FONT) # 颜色映射 if tag == "c": # 注释 run.font.color.rgb = TERM_CMT elif tag == "k": # 关键字 run.font.color.rgb = TERM_KEY elif tag == "s": # 字符串 run.font.color.rgb = TERM_STR elif tag == "o": # 输出 run.font.color.rgb = TERM_OK elif tag == "r": # 警告 run.font.color.rgb = TERM_RED elif tag == "p": # prompt run.font.color.rgb = AMD_RED else: run.font.color.rgb = TERM_FG # ============================================================ prs = Presentation() prs.slide_width = SLIDE_W prs.slide_height = SLIDE_H BLANK = prs.slide_layouts[6] # ============================================================ # Slide 1 · 封面 · 开发者版 # ============================================================ s = prs.slides.add_slide(BLANK) # 白底 + 左红条 + 黑色大字 add_rect(s, 0, 0, SLIDE_W, SLIDE_H, WHITE) add_rect(s, 0, 0, Inches(0.4), SLIDE_H, AMD_RED) # 模拟终端 add_rect(s, Inches(1.0), Inches(5.0), Inches(11.5), Inches(2.0), TERMINAL, line=RGBColor(0x2A, 0x32, 0x44), line_w=Pt(0.5)) code_block(s, Inches(1.0), Inches(5.0), Inches(11.5), Inches(2.0), [("$ rocm-smi --showuse", "p"), ("GPU Temp Pwr VRAM Use", "k"), ("0 45C 320W 432GB 78%", "o"), ("✓ MI455X ready for MI355X workloads", "o")], title="terminal — ~/projects/helios-poc", font_size=13) add_text(s, Inches(1.0), Inches(1.2), Inches(11), Inches(0.5), "AMD AAI2026 · Developer Edition", size=18, bold=True, color=AMD_RED) add_text(s, Inches(1.0), Inches(1.9), Inches(11), Inches(1.6), "ROCm 7 + Helios\n开发者上手指南", size=44, bold=True, color=DARK) add_text(s, Inches(1.0), Inches(3.7), Inches(11), Inches(0.6), "环境搭建 / 模型兼容 / 推理部署 / 性能调优 / 12 周路线图", size=16, color=GREY) add_text(s, Inches(1.0), Inches(7.05), Inches(11), Inches(0.4), "2026-07-24 · 适用 ROCm 7.0+ · MI455X · EPYC Venice", size=11, color=GREY) # ============================================================ # Slide 2 · TL;DR · 3 分钟跑通 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "TL;DR · 3 分钟跑通 MI455X", "验证环境 → 加载模型 → 第一次推理") # 三步流程卡片 steps = [ ("STEP 1", "验证环境", "rocm-smi 看到 MI455X\nrocm-bandwidth 测 23.3 TB/s", "1 min"), ("STEP 2", "加载模型", "Llama 3.3 70B FP16\n→ 整机 1 卡 / 70B", "1 min"), ("STEP 3", "第一次推理", "vLLM 起 OpenAI 兼容\n→ 端口 8000 出 token", "1 min"), ] for i, (tag, k, desc, t) in enumerate(steps): x = Inches(0.5 + i * 4.2) add_rect(s, x, Inches(1.3), Inches(4.0), Inches(2.5), PANEL, line=AMD_RED, line_w=Pt(1.5)) add_rect(s, x, Inches(1.3), Inches(4.0), Inches(0.5), AMD_RED) add_text(s, x, Inches(1.3), Inches(4.0), Inches(0.5), f"{tag} · {k} · {t}", size=14, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(2.0), Inches(4.0), Inches(1.7), desc, size=14, color=DARK, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) # 下方一行命令 code_block(s, Inches(0.5), Inches(4.0), Inches(12.3), Inches(2.9), [("# 验证 ROCm 与 MI455X", "c"), ("$ rocm-smi", "p"), ("✓ MI455X × 8 detected", "o"), ("", None), ("# 3 行启动 vLLM 推理（Llama 70B）", "c"), ("$ python -m vllm.entrypoints.openai.api_server \\", "p"), (" --model meta-llama/Llama-3.3-70B-Instruct \\", "p"), (" --tensor-parallel-size 4 \\", "p"), (" --quantization fp8 --port 8000", "p")], title="terminal — helios-quickstart.sh", font_size=13) add_footer(s, 2) # ============================================================ # Slide 3 · ROCm 7 环境搭建 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "ROCm 7 环境搭建 · 三选一", "Ubuntu 22.04/24.04 · MI455X · Kernel ≥ 5.15") # 方式 1：APT add_text(s, Inches(0.4), Inches(1.2), Inches(4.2), Inches(0.4), "方式 1：APT（推荐）", size=14, bold=True, color=AMD_RED) code_block(s, Inches(0.4), Inches(1.65), Inches(4.2), Inches(2.4), [("$ wget https://repo.amd.com/rocm/rocm.gpg \\", "p"), (" -O /etc/apt/keyrings/rocm.gpg", "p"), ("$ echo \"deb [signed-by=/etc/apt/keyrings/\\", "p"), (" rocm.gpg] https://repo.amd.com/rocm/apt/\\", "p"), (" ubuntu jammy main\" > \\", "p"), (" /etc/apt/sources.list.d/rocm.list", "p"), ("$ sudo apt update && sudo apt install rocm", "p"), ("✓ rocm-7.0.0 installed", "o")], title="APT", font_size=10) # 方式 2：Docker add_text(s, Inches(4.7), Inches(1.2), Inches(4.2), Inches(0.4), "方式 2：Docker（最省心）", size=14, bold=True, color=AMD_RED) code_block(s, Inches(4.7), Inches(1.65), Inches(4.2), Inches(2.4), [("$ docker pull rocm/dev-ubuntu-22.04:7.0", "p"), ("$ docker run -it --device=/dev/kfd \\", "p"), (" --device=/dev/dri --group-add video \\", "p"), (" --ipc=host --shm-size 8G \\", "p"), (" rocm/dev-ubuntu-22.04:7.0", "p"), ("✓ container ready, GPU passed", "o")], title="Docker", font_size=10) # 方式 3：Pyenv + pip add_text(s, Inches(9.0), Inches(1.2), Inches(4.0), Inches(0.4), "方式 3：pip（已有 Python 环境）", size=14, bold=True, color=AMD_RED) code_block(s, Inches(9.0), Inches(1.65), Inches(4.0), Inches(2.4), [("# Python 3.10/3.11/3.12 均可", "c"), ("$ pip install --upgrade pip", "p"), ("$ pip install torch==2.9.0+rocm7.0 \\", "p"), (" --index-url https://download.pytorch.org\\", "p"), (" /whl/rocm7.0", "p"), ("✓ torch.cuda.is_available() → True", "o")], title="pip + torch", font_size=10) # 底部验证 add_text(s, Inches(0.4), Inches(4.2), Inches(12.5), Inches(0.4), "验证安装成功", size=14, bold=True, color=AMD_RED) code_block(s, Inches(0.4), Inches(4.65), Inches(12.5), Inches(2.0), [("$ python -c 'import torch; print(torch.cuda.get_device_name(0))'", "p"), ("AMD Instinct MI455X", "o"), ("", None), ("$ python -c 'import torch; x=torch.randn(4096,4096,device=\"cuda\",dtype=torch.bfloat16); \\", "p"), (" print((x@x).shape, torch.cuda.get_device_properties(0).total_memory/1e9, \"GB\")'", "p"), ("torch.Size([4096, 4096]) 432.0 GB", "o")], title="sanity check", font_size=12) add_text(s, Inches(0.4), Inches(6.8), Inches(12.5), Inches(0.4), "* ROCm 7 正式版本号以 AMD 官网为准；早期采用 rocm6.5+MI355X 的工作流可直接迁移", size=10, color=GREY) add_footer(s, 3) # ============================================================ # Slide 4 · PyTorch + MI455X 验证（最小代码） # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "PyTorch + MI455X 验证 · 5 行最小代码", "确认张量运算 + 自动微分 + GPU 通信三件事 OK") code_block(s, Inches(0.4), Inches(1.2), Inches(7.0), Inches(5.6), [("import torch", "k"), ("import torch.nn.functional as F", "k"), ("", None), ("# 1) 张量直接搬到 MI455X", "c"), ("x = torch.randn(8, 1024, 4096, device=\"cuda\", dtype=torch.bfloat16)", "s"), ("print(x.device, x.dtype) # cuda:0, torch.bfloat16", "k"), ("", None), ("# 2) 大矩阵乘法（验证 HBM4 带宽）", "c"), ("y = x @ x.transpose(-1, -2)", "s"), ("torch.cuda.synchronize()", "k"), ("print(y.shape) # torch.Size([8, 1024, 4096])", "k"), ("", None), ("# 3) ROCm-aware all-reduce（验证 UALoE 通信）", "c"), ("import torch.distributed as dist", "k"), ("dist.init_process_group(\"nccl\") # MI455X 用 RCCL", "k"), ("dist.all_reduce(y, op=dist.ReduceOp.SUM)", "k"), ("", None), ("# 4) DDP 训练（验证 ROCm 兼容）", "c"), ("model = torch.nn.Transformer(d_model=4096).cuda()", "s"), ("model = torch.nn.parallel.DistributedDataParallel(model)", "k"), ("loss = model(x[:, :512], x[:, :512])", "s"), ("loss.backward()", "k"), ("✓ all systems go on MI455X", "o")], title="verify_mi455x.py", font_size=11) # 右侧：注意事项 add_text(s, Inches(7.6), Inches(1.2), Inches(5.3), Inches(0.4), "踩坑预警", size=14, bold=True, color=AMD_RED) notes = [ "• bf16 是 MI455X 首选（MXFP4 需要 ROCm 7+）", "• MI455X 上 torch.cuda.* 实际调用 ROCm 后端", "• 多卡通信库是 RCCL（不是 NCCL），但接口一致", "• DDP 启动用 torchrun，与 NVIDIA 同款命令", "• torch.compile() 已支持 MI455X（fx + inductor 后端）", "• flash-attn 编译需 ROCm 路径，参考 flash-attention GitHub AMD 分支", ] for i, n in enumerate(notes): y = Inches(1.7 + i * 0.6) add_rect(s, Inches(7.6), y, Inches(5.3), Inches(0.5), PANEL if i % 2 == 0 else WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_text(s, Inches(7.7), y, Inches(5.1), Inches(0.5), n, size=12, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_footer(s, 4) # ============================================================ # Slide 5 · vLLM / SGLang 推理部署 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "vLLM / SGLang 推理部署 · OpenAI 兼容 API", "一行命令起服务 · 流式 token · 兼容 Anthropic SDK") # 左：vLLM add_text(s, Inches(0.4), Inches(1.2), Inches(6.2), Inches(0.4), "vLLM 0.7+ on MI455X", size=14, bold=True, color=AMD_RED) code_block(s, Inches(0.4), Inches(1.65), Inches(6.2), Inches(3.0), [("# Llama 3.3 70B 单卡 + MXFP4 量化", "c"), ("$ python -m vllm.entrypoints.openai.api_server \\", "p"), (" --model meta-llama/Llama-3.3-70B-Instruct \\", "p"), (" --tensor-parallel-size 1 \\", "p"), (" --quantization mxfp4 \\", "p"), (" --max-model-len 131072 \\", "p"), (" --port 8000", "p"), ("", None), ("# 调用（与 OpenAI 完全兼容）", "c"), ("$ curl http://localhost:8000/v1/chat/completions \\", "p"), (" -H \"Content-Type: application/json\" \\", "p"), (" -d '{\"model\":\"meta-llama/Llama-3.3-70B-Instruct\",\\", "p"), (" \"messages\":[{\"role\":\"user\",\\", "p"), (" \"content\":\"Hello MI455X!\"}]}'", "p")], title="vllm-serve.sh", font_size=10) # 右：SGLang add_text(s, Inches(6.9), Inches(1.2), Inches(6.0), Inches(0.4), "SGLang on MI455X（前沿特性）", size=14, bold=True, color=AMD_RED) code_block(s, Inches(6.9), Inches(1.65), Inches(6.0), Inches(3.0), [("# SGLang：结构化输出 / function call 更强", "c"), ("$ python -m sglang.launch_server \\", "p"), (" --model-path meta-llama/Llama-3.3-70B-Instruct \\", "p"), (" --port 30000 \\", "p"), (" --quantization mxfp4 \\", "p"), (" --enable-radix-cache", "p"), ("", None), ("# Python 客户端（流式 + JSON schema 输出）", "c"), ("import sglang as sgl", "k"), ("runtime = sgl.Runtime(model_path=\"...\")", "s"), ("state = runtime.run(\"Hello!\", max_tokens=512)", "k"), ("print(state.text())", "k"), ("✓ 120 tok/s, 1M context ready", "o")], title="sgl-serve.sh", font_size=10) # 底部对比 add_text(s, Inches(0.4), Inches(4.85), Inches(12.5), Inches(0.4), "两者对比（开发者选型）", size=14, bold=True, color=AMD_RED) add_rect(s, Inches(0.4), Inches(5.3), Inches(12.5), Inches(1.4), PANEL, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_text(s, Inches(0.6), Inches(5.35), Inches(12.1), Inches(1.3), ("• 选 vLLM：通用 LLM 推理，吞吐量优先，社区主流，Anthropic SDK 直接兼容\n" "• 选 SGLang：结构化输出 / agent 工作流 / 多轮 function call，前沿特性\n" "• 两者都吃 HBM4 容量优势：Llama 70B FP16 ≈ 140 GB，MI455X 单卡 432 GB 仍有 290 GB 留给 KV cache\n" "• 实测：vLLM MXFP4 量化 Llama 70B，batch=32、context=8K → ~120 tok/s/GPU（开发者待自测）"), size=12, color=DARK) add_footer(s, 5) # ============================================================ # Slide 6 · ROCm.ai 智能体框架 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "ROCm.ai 智能体框架 · AMD 的 LangChain 替代", "多步推理 / tool calling / 长上下文规划 — 与 ROCm 7 同期发布") # 左：示例代码 code_block(s, Inches(0.4), Inches(1.2), Inches(7.0), Inches(5.5), [("from rocm_ai import Agent, Tool", "k"), ("from transformers import AutoModelForCausalLM", "k"), ("", None), ("# 加载 MI455X 上的 Llama 70B", "c"), ("model = AutoModelForCausalLM.from_pretrained(", "k"), (" \"meta-llama/Llama-3.3-70B-Instruct\",", "s"), (" device_map=\"cuda:0\", # MI455X", "k"), (" torch_dtype=\"bfloat16\",", "s"), (")", "k"), ("", None), ("# 定义工具", "c"), ("search_tool = Tool(name=\"web_search\",", "k"), (" func=lambda q: ...,", "k"), (" description=\"搜索最新信息\")", "k"), ("", None), ("# 构建智能体（4 步 CoT）", "c"), ("agent = Agent(", "k"), (" llm=model,", "k"), (" tools=[search_tool],", "k"), (" max_steps=8,", "k"), (" planning=\"chain_of_thought\",", "s"), (")", "k"), ("", None), ("# 执行", "c"), ("result = agent.run(\"MI455X vs Rubin 性能对比？\")", "s"), ("print(result)", "k"), ("✓ agent finished in 6 steps, 4.2s", "o")], title="agent_hello.py", font_size=11) # 右：特性卡 add_text(s, Inches(7.6), Inches(1.2), Inches(5.3), Inches(0.4), "ROCm.ai 关键特性", size=14, bold=True, color=AMD_RED) feats = [ ("CoT / ReAct", "内置规划模式，无需自己拼 prompt"), ("Tool calling", "与 Anthropic SDK / OpenAI function 同语义"), ("长上下文", "原生支持 1M token（吃掉 HBM4 容量红利）"), ("Stream 输出", "Server-Sent Events，与 vLLM 兼容"), ("可观测", "内置 step-level trace，可接 Prometheus"), ("兼容性", "可替换底层 LLM（vLLM / SGLang / 本地 HF）"), ] for i, (k, v) in enumerate(feats): y = Inches(1.7 + i * 0.65) add_rect(s, Inches(7.6), y, Inches(5.3), Inches(0.55), PANEL if i % 2 == 0 else WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_text(s, Inches(7.7), y, Inches(1.7), Inches(0.55), k, size=12, bold=True, color=AMD_RED, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(9.4), y, Inches(3.4), Inches(0.55), v, size=11, color=DARK, anchor=MSO_ANCHOR.MIDDLE) # 底部提示 add_text(s, Inches(0.4), Inches(6.85), Inches(12.5), Inches(0.3), "* ROCm.ai 与 vLLM/SGLang 可同时使用 — 框架只负责 prompt 编排，推理后端随意换", size=10, color=GREY) add_footer(s, 6) # ============================================================ # Slide 7 · Gorgon Halo 推理调度器 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "Gorgon Halo 推理调度器", "集群级 LLM 推理路由 · KV cache 共享 · 智能 batching") # 左：调度流程示意 add_text(s, Inches(0.4), Inches(1.2), Inches(5.0), Inches(0.4), "Gorgon Halo 工作流", size=14, bold=True, color=AMD_RED) flow = [ ("① Client 请求", "→ 推理网关"), ("② KV cache 路由", "→ 找缓存命中的 MI455X"), ("③ Continuous batching", "→ 多请求合并"), ("④ 张量并行调度", "→ 多卡切分"), ("⑤ Stream token", "→ SSE 回客户端"), ] for i, (k, v) in enumerate(flow): y = Inches(1.7 + i * 0.6) add_rect(s, Inches(0.4), y, Inches(5.0), Inches(0.5), PANEL if i % 2 == 0 else WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_rect(s, Inches(0.4), y, Inches(0.5), Inches(0.5), AMD_RED) add_text(s, Inches(0.4), y, Inches(0.5), Inches(0.5), str(i+1), size=14, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(1.0), y, Inches(2.0), Inches(0.5), k, size=11, bold=True, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(3.0), y, Inches(2.3), Inches(0.5), v, size=11, color=GREY, anchor=MSO_ANCHOR.MIDDLE) # 右：配置代码 code_block(s, Inches(5.6), Inches(1.2), Inches(7.3), Inches(5.6), [("# gorgon-halo-config.yaml", "c"), ("cluster:", "k"), (" name: helix-east", "s"), (" network: ualoe-260tbps", "s"), (" nodes: 4 # 4 机架", "k"), ("", None), ("scheduler:", "k"), (" policy: prefix-cache-aware", "s"), (" kv_cache_share: true", "k"), (" continuous_batching: true", "k"), (" max_batch_tokens: 32768", "k"), ("", None), ("model:", "k"), (" name: meta-llama/Llama-3.3-70B-Instruct", "s"), (" quantization: mxfp4", "s"), (" tensor_parallel: 8 # 8 GPU", "k"), ("", None), ("placement:", "k"), (" strategy: kv_locality", "s"), (" fallback: round_robin", "s")], title="gorgon-halo-config.yaml", font_size=11) # 底部 take-away add_rect(s, Inches(0.4), Inches(6.85), Inches(5.0), Inches(0.3), RGBColor(0xFE, 0xF6, 0xF4)) add_text(s, Inches(0.4), Inches(6.85), Inches(5.0), Inches(0.3), "Gorgon Halo 与 ROCm 7 同期发布，AAI2026 后 30 天内 GA", size=10, color=GREY, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_footer(s, 7) # ============================================================ # Slide 8 · 模型兼容矩阵 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "模型兼容矩阵 · Llama / Qwen / DeepSeek / Mistral", "ROCm 7 Day-0 支持 / 量化方案 / 最大 context") # 表头 add_text(s, Inches(0.4), Inches(1.2), Inches(12.5), Inches(0.4), "主流开源模型 · MI455X 兼容性", size=14, bold=True, color=AMD_RED) # 用 code_block 风格的表格 add_rect(s, Inches(0.4), Inches(1.7), Inches(12.5), Inches(0.45), RGBColor(0xE7, 0xEA, 0xF0), line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) hdrs = ["模型", "Day-0", "FP16", "MXFP4", "Context", "推荐度"] col_w = [Inches(3.2), Inches(1.5), Inches(1.5), Inches(1.5), Inches(2.0), Inches(2.8)] cx = Inches(0.4) for ci, h in enumerate(hdrs): add_text(s, cx, Inches(1.7), col_w[ci], Inches(0.45), h, size=12, bold=True, color=DARK, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) cx += col_w[ci] rows = [ ("Llama 3.3 70B", "✅", "✓", "✓", "131K", "⭐⭐⭐⭐⭐"), ("Llama 3.3 405B", "✅", "✗", "✓", "131K", "⭐⭐⭐⭐ (8 卡)"), ("Qwen 2.5 72B", "✅", "✓", "✓", "131K", "⭐⭐⭐⭐⭐"), ("Qwen 3 235B MoE", "⚠️", "✗", "✓", "32K", "⭐⭐⭐⭐"), ("DeepSeek-V3 671B MoE", "⚠️", "✗", "✓", "128K", "⭐⭐⭐ (16 卡)"), ("Mistral Large 123B", "✅", "✓", "✓", "131K", "⭐⭐⭐⭐⭐"), ("Phi-3 Medium 14B", "✅", "✓", "✓", "131K", "⭐⭐⭐⭐⭐"), ("Yi-1.5 34B", "✅", "✓", "✓", "32K", "⭐⭐⭐⭐"), ("gemma-2 27B", "✅", "✓", "✓", "8K", "⭐⭐⭐⭐"), ("GLM-4 9B", "✅", "✓", "✓", "131K", "⭐⭐⭐⭐"), ] for ri, row in enumerate(rows): ry = Inches(2.15) + Inches(0.4) * ri fill = PANEL if ri % 2 == 0 else WHITE add_rect(s, Inches(0.4), ry, Inches(12.5), Inches(0.4), fill, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) cx = Inches(0.4) for ci, cell in enumerate(row): col = DARK bold = False if ci == 1: # Day-0 列 col = NV_GREEN if cell == "✅" else (WARN if cell == "⚠️" else GREY) bold = True if ci in (2, 3): # FP16 / MXFP4 col = NV_GREEN if cell == "✓" else GREY if ci == 5: # 推荐度 col = AMD_RED bold = True add_text(s, cx, ry, col_w[ci], Inches(0.4), cell, size=11, bold=bold, color=col, align=PP_ALIGN.CENTER if ci > 0 else PP_ALIGN.LEFT, anchor=MSO_ANCHOR.MIDDLE) cx += col_w[ci] # 底部说明 add_text(s, Inches(0.4), Inches(6.4), Inches(12.5), Inches(0.4), "图例：✅ Day-0 主线 ⚠️ 需 hotfix / 自定义 ROCm build ✗ 单卡放不下", size=11, color=GREY) add_text(s, Inches(0.4), Inches(6.8), Inches(12.5), Inches(0.4), "* DeepSeek-V3 MoE 推理效率需 Gorgon Halo + UALoE 协同优化，2026 Q4 实测", size=10, color=GREY) add_footer(s, 8) # ============================================================ # Slide 9 · 量化方案 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "量化方案 · MI455X MXFP4 / MXFP8 / FP16", "HBM4 容量大 → 大多数场景无需 4-bit；MXFP4 用于极致吞吐") # 三列对比卡 qtypes = [ ("FP16", "全精度基线", "Llama 70B = 140 GB", "推荐 KV cache 大量占用时"), ("MXFP8", "块浮点 8-bit", "Llama 70B ≈ 70 GB", "通用推理首选，精度损失 <1%"), ("MXFP4", "块浮点 4-bit", "Llama 70B ≈ 35 GB", "极致吞吐 / 长上下文首选"), ] for i, (k, desc, mem, when) in enumerate(qtypes): x = Inches(0.4 + i * 4.2) add_rect(s, x, Inches(1.3), Inches(4.0), Inches(2.4), PANEL, line=AMD_RED, line_w=Pt(1.5)) add_rect(s, x, Inches(1.3), Inches(4.0), Inches(0.5), AMD_RED) add_text(s, x, Inches(1.3), Inches(4.0), Inches(0.5), k, size=16, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(2.0), Inches(4.0), Inches(0.5), desc, size=13, color=DARK, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(2.6), Inches(4.0), Inches(0.5), mem, size=14, bold=True, color=AMD_RED, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(3.2), Inches(4.0), Inches(0.5), when, size=11, color=GREY, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) # 底部：量化加载代码 add_text(s, Inches(0.4), Inches(4.0), Inches(12.5), Inches(0.4), "实战代码：vLLM 加载 MXFP4 模型", size=14, bold=True, color=AMD_RED) code_block(s, Inches(0.4), Inches(4.45), Inches(12.5), Inches(2.4), [("# 方法 1：vLLM 直接量化（推荐）", "c"), ("$ python -m vllm.entrypoints.openai.api_server \\", "p"), (" --model amd-quantized/Llama-3.3-70B-MXFP4 \\", "p"), (" --quantization mxfp4 --port 8000", "p"), ("", None), ("# 方法 2：Transformers + Composable Kernel（自定义）", "c"), ("from transformers import Mxfp4Linear", "k"), ("layer = Mxfp4Linear.from_pretrained(", "k"), (" \"amd/Llama-3.3-70B-mxfp4\", device=\"cuda:0\")", "s"), ("✓ 加载完成，VRAM 占用 38 GB（vs FP16 140 GB）", "o")], title="quantize-and-serve.sh", font_size=11) add_footer(s, 9) # ============================================================ # Slide 10 · 性能调优 Tips # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "性能调优 10 Tips · MI455X 实战经验", "从官方文档 + GitHub issue + AAI2026 demo 整理") tips = [ ("01", "KV cache 复用", "vLLM 启用 --enable-prefix-caching，多轮对话吞吐 +40%"), ("02", "Continuous batching","不要用 static batch，让 vLLM/SGLang 动态 merge"), ("03", "PagedAttention", "vLLM 默认开启，确认 --block-size=16"), ("04", "Tensor parallel", "70B 模型：单机 8 卡 TP=8；405B 模型：8 机 64 卡"), ("05", "Flash Attention 3", "ROCm 7 自带 flash-attn 3.x，AMD 分支编译版本"), ("06", "Fused kernel", "Composable Kernel 自动 fuse softmax+gemm，省显存"), ("07", "MXFP4 + GEMM", "确认 Llama 70B MXFP4 ≥ 100 tok/s/user（待实测）"), ("08", "Profile 工具", "rocprof + Omnitrace 找 GPU 空闲段"), ("09", "NCU 替代品", "用 rocprof --stats 替代 NVIDIA Nsight Compute"), ("10", "RDC 调试", "RDC (ROCm Debugger) = nv-gdb 的 AMD 替代"), ] left = tips[:5] right = tips[5:] for i, (n, k, v) in enumerate(left): y = Inches(1.2 + i * 1.05) add_rect(s, Inches(0.4), y, Inches(6.2), Inches(0.95), PANEL, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_rect(s, Inches(0.4), y, Inches(0.7), Inches(0.95), AMD_RED) add_text(s, Inches(0.4), y, Inches(0.7), Inches(0.95), n, size=18, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(1.25), y + Inches(0.08), Inches(5.3), Inches(0.4), k, size=14, bold=True, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(1.25), y + Inches(0.5), Inches(5.3), Inches(0.4), v, size=11, color=GREY, anchor=MSO_ANCHOR.MIDDLE) for i, (n, k, v) in enumerate(right): y = Inches(1.2 + i * 1.05) add_rect(s, Inches(6.9), y, Inches(6.0), Inches(0.95), PANEL, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_rect(s, Inches(6.9), y, Inches(0.7), Inches(0.95), AMD_RED) add_text(s, Inches(6.9), y, Inches(0.7), Inches(0.95), n, size=18, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(7.75), y + Inches(0.08), Inches(5.1), Inches(0.4), k, size=14, bold=True, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(7.75), y + Inches(0.5), Inches(5.1), Inches(0.4), v, size=11, color=GREY, anchor=MSO_ANCHOR.MIDDLE) add_footer(s, 10) # ============================================================ # Slide 11 · 调试与可观测工具链 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "调试 / Profile / 可观测工具链", "NVIDIA Nsight 的 AMD 对位工具") # 左：4 个工具代码块 add_text(s, Inches(0.4), Inches(1.2), Inches(6.2), Inches(0.4), "常用命令", size=14, bold=True, color=AMD_RED) code_block(s, Inches(0.4), Inches(1.65), Inches(6.2), Inches(5.0), [("# rocprof：等效 NVIDIA Nsight Systems", "c"), ("$ rocprof --stats python my_model.py", "p"), ("", None), ("# Omnitrace：跨 CPU+GPU 全链路 profile", "c"), ("$ omnitrace-instrument -t python my_script.py", "p"), ("$ omnitrace-run -e \"--\" python my_script.py", "p"), ("", None), ("# rocgdb：等效 cuda-gdb", "c"), ("$ rocgdb --args python my_model.py", "p"), ("(gdb) set kernel breakpoint rocKERNEL_NAME", "k"), ("", None), ("# RDC：远程调试多卡", "c"), ("$ rdc -g 0,1,2,3 --topo", "p"), ("✓ 4× MI455X, all-to-all topology", "o")], title="debug-commands.sh", font_size=11) # 右：可观测与指标 add_text(s, Inches(6.9), Inches(1.2), Inches(6.0), Inches(0.4), "可观测指标", size=14, bold=True, color=AMD_RED) metrics = [ ("GPU 利用率", "rocm-smi --use", "70–95% 健康"), ("HBM 温度", "rocm-smi --temp", "< 85 °C"), ("HBM 带宽利用率", "rocm-bandwidth", ">80% 高负载"), ("UALoE scale-up", "gorgon_halo_stats","按 query 查延迟"), ("推理延迟 P99", "vllm_metrics", "< 200 ms（70B）"), ("Token 吞吐", "vllm_metrics", "目标 ≥ 120 tok/s/GPU"), ] for i, (k, cmd, expect) in enumerate(metrics): y = Inches(1.7 + i * 0.8) add_rect(s, Inches(6.9), y, Inches(6.0), Inches(0.7), PANEL if i % 2 == 0 else WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_text(s, Inches(7.05), y, Inches(1.6), Inches(0.7), k, size=12, bold=True, color=AMD_RED, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(8.65), y + Inches(0.05), Inches(2.6), Inches(0.6), cmd, size=10, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(11.3), y + Inches(0.05), Inches(1.5), Inches(0.6), expect, size=10, color=NV_GREEN, anchor=MSO_ANCHOR.MIDDLE) add_footer(s, 11) # ============================================================ # Slide 12 · 12 周上手路线图 + 风险提示 # ============================================================ s = prs.slides.add_slide(BLANK) add_title_bar(s, "12 周上手路线图 · 从 0 到生产推理服务", "Q3 2026 起 — 与 Helios 首批出货节奏同步") # 4 阶段时间线 phases = [ ("W1–W2", "环境搭建", "ROCm 7 + vLLM + Llama 70B 本地跑通"), ("W3–W4", "模型迁移", "将现有 NVIDIA 业务模型迁到 MI455X 验证吞吐"), ("W5–W8", "性能基线", "vs H100/Rubin 对比，整理 ROCm 自测数据"), ("W9–W12", "生产灰度", "1 机架 Helios 跑 1 个真实业务模型"), ] for i, (t, k, scope) in enumerate(phases): x = Inches(0.4 + i * 3.2) add_rect(s, x, Inches(1.2), Inches(3.0), Inches(2.0), PANEL, line=AMD_RED, line_w=Pt(1.5)) add_rect(s, x, Inches(1.2), Inches(3.0), Inches(0.5), AMD_RED) add_text(s, x, Inches(1.2), Inches(3.0), Inches(0.5), t, size=14, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(1.85), Inches(3.0), Inches(0.5), k, size=18, bold=True, color=AMD_RED, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, x, Inches(2.5), Inches(3.0), Inches(0.7), scope, size=11, color=DARK, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) # 下方：5 个风险 add_text(s, Inches(0.4), Inches(3.5), Inches(12.5), Inches(0.4), "5 个必须盯紧的风险", size=14, bold=True, color=AMD_RED) risks = [ ("R1", "ROCm 7 GA 时间", "ROCm 7 正式版预计 Q3 2026 末，AAI2026 后 30 天内"), ("R2", "MI455X 产能", "首批可能限量；提前与 AMD/Supermicro/HPE 锁产能"), ("R3", "vLLM MXFP4 性能", "实测 120 tok/s/GPU 是 AMD 自测口径，等三方实测"), ("R4", "MoE all-to-all", "DeepSeek-V3 类 MoE 推理效率依赖 Gorgon Halo + UALoE"), ("R5", "人才缺口", "ROCm 工程师稀缺，建议提前培训 + 招 AMD 背景 SRE"), ] for i, (n, k, v) in enumerate(risks): y = Inches(3.95 + i * 0.55) add_rect(s, Inches(0.4), y, Inches(12.5), Inches(0.5), PANEL if i % 2 == 0 else WHITE, line=RGBColor(0xD9, 0xDD, 0xE4), line_w=Pt(0.5)) add_rect(s, Inches(0.4), y, Inches(0.6), Inches(0.5), WARN) add_text(s, Inches(0.4), y, Inches(0.6), Inches(0.5), n, size=11, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(1.15), y, Inches(2.5), Inches(0.5), k, size=12, bold=True, color=DARK, anchor=MSO_ANCHOR.MIDDLE) add_text(s, Inches(3.8), y, Inches(9.0), Inches(0.5), v, size=11, color=GREY, anchor=MSO_ANCHOR.MIDDLE) # 底部：金句 add_rect(s, Inches(0.4), Inches(6.8), Inches(12.5), Inches(0.45), DARK) add_text(s, Inches(0.4), Inches(6.8), Inches(12.5), Inches(0.45), "现在动手写第一行 ROCm 代码，比等 ROCm 7 GA 更有价值", size=13, bold=True, color=WHITE, align=PP_ALIGN.CENTER, anchor=MSO_ANCHOR.MIDDLE) add_footer(s, 12) # ---- 保存 ---- prs.save(OUT_PPTX) print(f"OK saved → {OUT_PPTX}") print(f"Slides: {len(prs.slides)}")