Skip to content

C
coderai
Commit 087ba9e1 authored by Stefy Lanza (nextime / spora )'s avatar Stefy Lanza (nextime / spora )
Browse files
Options

Initial commit: Add CoderAI OpenAI-compatible API server 

- Add main server script with FastAPI and memory-aware model loading
- Add requirements.txt with dependencies and platform-specific PyTorch options
- Add comprehensive README.md with installation, usage, and troubleshooting
- Add LICENSE.md with GPLv3 license
parents
Pipeline #238 failed with stages
      Show whitespace changes
      Inline Side-by-side
      Showing with 2290 additions and 0 deletions
      LICENSE.md 0 → 100644
      View file @ 087ba9e1
      GNU GENERAL PUBLIC LICENSE
      Version 3, 29 June 2007
      Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
      Everyone is permitted to copy and distribute verbatim copies
      of this license document, but changing it is not allowed.
      Preamble
      The GNU General Public License is a free, copyleft license for
      software and other kinds of works.
      The licenses for most software and other practical works are designed
      to take away your freedom to share and change the works. By contrast,
      the GNU General Public License is intended to guarantee your freedom to
      share and change all versions of a program--to make sure it remains free
      software for all its users. We, the Free Software Foundation, use the
      GNU General Public License for most of our software; it applies also to
      any other work released this way by its authors. You can apply it to
      your programs, too.
      When we speak of free software, we are referring to freedom, not
      price. Our General Public Licenses are designed to make sure that you
      have the freedom to distribute copies of free software (and charge for
      them if you wish), that you receive source code or can get it if you
      want it, that you can change the software or use pieces of it in new
      free programs, and that you know you can do these things.
      To protect your rights, we need to prevent others from denying you
      these rights or asking you to surrender the rights. Therefore, you have
      certain responsibilities if you distribute copies of the software, or if
      you modify it: responsibilities to respect the freedom of others.
      For example, if you distribute copies of such a program, whether
      gratis or for a fee, you must pass on to the recipients the same
      freedoms that you received. You must make sure that they, too, receive
      or can get the source code. And you must show them these terms so they
      know their rights.
      Developers that use the GNU GPL protect your rights with two steps:
      (1) assert copyright on the software, and (2) offer you this License
      giving you legal permission to copy, distribute and/or modify it.
      For the developers' and authors' protection, the GPL clearly explains
      that there is no warranty for this free software. For both users' and
      authors' sake, the GPL requires that modified versions be marked as
      changed, so that their problems will not be attributed erroneously to
      authors of previous versions.
      Some devices are designed to deny users access to install or run
      modified versions of the software inside them, although the manufacturer
      can do so. This is fundamentally incompatible with the aim of
      protecting users' freedom to change the software. The systematic
      pattern of such abuse occurs in the area of products for individuals to
      use, which is precisely where it is most unacceptable. Therefore, we
      have designed this version of the GPL to prohibit the practice for those
      products. If such problems arise substantially in other domains, we
      stand ready to extend this provision to those domains in future versions
      of the GPL, as needed to protect the freedom of users.
      Finally, every program is threatened constantly by software patents.
      States should not allow patents to restrict development and use of
      software on general-purpose computers, but in those that do, we wish to
      avoid the special danger that patents applied to a free program could
      make it effectively proprietary. To prevent this, the GPL assures that
      patents cannot be used to render the program non-free.
      The precise terms and conditions for copying, distribution and
      modification follow.
      TERMS AND CONDITIONS
      0. Definitions.
      "This License" refers to version 3 of the GNU General Public License.
      "Copyright" also means copyright-like laws that apply to other kinds of
      works, such as semiconductor masks.
      "The Program" refers to any copyrightable work licensed under this
      License. Each licensee is addressed as "you". "Licensees" and
      "recipients" may be individuals or organizations.
      To "modify" a work means to copy from or adapt all or part of the work
      in a fashion requiring copyright permission, other than the making of an
      exact copy. The resulting work is called a "modified version" of the
      earlier work or a work "based on" the earlier work.
      A "covered work" means either the unmodified Program or a work based
      on the Program.
      To "propagate" a work means to do anything with it that, without
      permission, would make you directly or secondarily liable for
      infringement under applicable copyright law, except executing it on a
      computer or modifying a private copy. Propagation includes copying,
      distribution (with or without modification), making available to the
      public, and in some countries other activities as well.
      To "convey" a work means any kind of propagation that enables other
      parties to make or receive copies. Mere interaction with a user through
      a computer network, with no transfer of a copy, is not conveying.
      An interactive user interface displays "Appropriate Legal Notices"
      to the extent that it includes a convenient and prominently visible
      feature that (1) displays an appropriate copyright notice, and (2)
      tells the user that there is no warranty for the work (except to the
      extent that warranties are provided), that licensees may convey the
      work under this License, and how to view a copy of this License. If
      the interface presents a list of user commands or options, such as a
      menu, a prominent item in the list meets this criterion.
      1. Source Code.
      The "source code" for a work means the preferred form of the work
      for making modifications to it. "Object code" means any non-source
      form of a work.
      A "Standard Interface" means an interface that either is an official
      standard defined by a recognized standards body, or, in the case of
      interfaces specified for a particular programming language, one that
      is widely used among developers working in that language.
      The "System Libraries" of an executable work include anything, other
      than the work as a whole, that (a) is included in the normal form of
      packaging a Major Component, but which is not part of that Major
      Component, and (b) serves only to enable use of the work with that
      Major Component, or to implement a Standard Interface for which an
      implementation is available to the public in source code form. A
      "Major Component", in this context, means a major essential component
      (kernel, window system, and so on) of the specific operating system
      (if any) on which the executable work runs, or a compiler used to
      produce the work, or an object code interpreter used to run it.
      The "Corresponding Source" for a work in object code form means all
      the source code needed to generate, install, and (for an executable
      work) run the object code and to modify the work, including scripts to
      control those activities. However, it does not include the work's
      System Libraries, or general-purpose tools or generally available free
      programs which are used unmodified in performing those activities but
      which are not part of the work. For example, Corresponding Source
      includes interface definition files associated with source files for
      the work, and the source code for shared libraries and dynamically
      linked subprograms that the work is specifically designed to require,
      such as by intimate data communication or control flow between those
      subprograms and other parts of the work.
      The Corresponding Source need not include anything that users
      can regenerate automatically from other parts of the Corresponding
      Source.
      The Corresponding Source for a work in source code form is that
      same work.
      2. Basic Permissions.
      All rights granted under this License are granted for the term of
      copyright on the Program, and are irrevocable provided the stated
      conditions are met. This License explicitly affirms your unlimited
      permission to run the unmodified Program. The output from running a
      covered work is covered by this License only if the output, given its
      content, constitutes a covered work. This License acknowledges your
      rights of fair use or other equivalent, as provided by copyright law.
      You may make, run and propagate covered works that you do not
      convey, without conditions so long as your license otherwise remains
      in force. You may convey covered works to others for the sole purpose
      of having them make modifications exclusively for you, or provide you
      with facilities for running those works, provided that you comply with
      the terms of this License in conveying all material for which you do
      not control copyright. Those thus making or running the covered works
      for you must do so exclusively on your behalf, under your direction
      and control, on terms that prohibit them from making any copies of
      your copyrighted material outside their relationship with you.
      Conveying under any other circumstances is permitted solely under
      the conditions stated below. Sublicensing is not allowed; section 10
      makes it unnecessary.
      3. Protecting Users' Legal Rights From Anti-Circumvention Law.
      No covered work shall be deemed part of an effective technological
      measure under any applicable law fulfilling obligations under article
      11 of the WIPO copyright treaty adopted on 20 December 1996, or
      similar laws prohibiting or restricting circumvention of such
      measures.
      When you convey a covered work, you waive any legal power to forbid
      circumvention of technological measures to the extent such circumvention
      is effected by exercising rights under this License with respect to
      the covered work, and you disclaim any intention to limit operation or
      modification of the work as a means of enforcing, against the work's
      users, your or third parties' legal rights to forbid circumvention of
      technological measures.
      4. Conveying Verbatim Copies.
      You may convey verbatim copies of the Program's source code as you
      receive it, in any medium, provided that you conspicuously and
      appropriately publish on each copy an appropriate copyright notice;
      keep intact all notices stating that this License and any
      non-permissive terms added in accord with section 7 apply to the code;
      keep intact all notices of the absence of any warranty; and give all
      recipients a copy of this License along with the Program.
      You may charge any price or no price for each copy that you convey,
      and you may offer support or warranty protection for a fee.
      5. Conveying Modified Source Versions.
      You may convey a work based on the Program, or the modifications to
      produce it from the Program, in the form of source code under the
      terms of section 4, provided that you also meet all of these conditions:
      a) The work must carry prominent notices stating that you modified
      it, and giving a relevant date.
      b) The work must carry prominent notices stating that it is
      released under this License and any conditions added under section
      7. This requirement modifies the requirement in section 4 to
      "keep intact all notices".
      c) You must license the entire work, as a whole, under this
      License to anyone who comes into possession of a copy. This
      License will therefore apply, along with any applicable section 7
      additional terms, to the whole of the work, and all its parts,
      regardless of how they are packaged. This License gives no
      permission to license the work in any other way, but it does not
      invalidate such permission if you have separately received it.
      d) If the work has interactive user interfaces, each must display
      Appropriate Legal Notices; however, if the Program has interactive
      interfaces that do not display Appropriate Legal Notices, your
      work need not make them do so.
      A compilation of a covered work with other separate and independent
      works, which are not by their nature extensions of the covered work,
      and which are not combined with it such as to form a larger program,
      in or on a volume of a storage or distribution medium, is called an
      "aggregate" if the compilation and its resulting copyright are not
      used to limit the access or legal rights of the compilation's users
      beyond what the individual works permit. Inclusion of a covered work
      in an aggregate does not cause this License to apply to the other
      parts of the aggregate.
      6. Conveying Non-Source Forms.
      You may convey a covered work in object code form under the terms
      of sections 4 and 5, provided that you also convey the
      machine-readable Corresponding Source under the terms of this License,
      in one of these ways:
      a) Convey the object code in, or embodied in, a physical product
      (including a physical distribution medium), accompanied by the
      Corresponding Source fixed on a durable physical medium
      customarily used for software interchange.
      b) Convey the object code in, or embodied in, a physical product
      (including a physical distribution medium), accompanied by a
      written offer, valid for at least three years and valid for as
      long as you offer spare parts or customer support for that product
      model, to give anyone who possesses the object code either (1) a
      copy of the Corresponding Source for all the software in the
      product that is covered by this License, on a durable physical
      medium customarily used for software interchange, for a price no
      more than your reasonable cost of physically performing this
      conveying of source, or (2) access to copy the
      Corresponding Source from a network server at no charge.
      c) Convey individual copies of the object code with a copy of the
      written offer to provide the Corresponding Source. This
      alternative is allowed only occasionally and noncommercially, and
      only if you received the object code with such an offer, in accord
      with subsection 6b.
      d) Convey the object code by offering access from a designated
      place (gratis or for a charge), and offer equivalent access to the
      Corresponding Source in the same way through the same place at no
      further charge. You need not require recipients to copy the
      Corresponding Source along with the object code. If the place to
      copy the object code is a network server, the Corresponding Source
      may be on a different server (operated by you or a third party)
      that supports equivalent copying facilities, provided you maintain
      clear directions next to the object code saying where to find the
      Corresponding Source. Regardless of what server hosts the
      Corresponding Source, you remain obligated to ensure that it is
      available for as long as needed to satisfy these requirements.
      e) Convey the object code using peer-to-peer transmission, provided
      you inform other peers where the object code and Corresponding
      Source of the work are being offered to the general public at no
      charge under subsection 6d.
      A separable portion of the object code, whose source code is excluded
      from the Corresponding Source as a System Library, need not be
      included in conveying the object code work.
      A "User Product" is either (1) a "consumer product", which means any
      tangible personal property which is normally used for personal, family,
      or household purposes, or (2) anything designed or sold for incorporation
      into a dwelling. In determining whether a product is a consumer product,
      doubtful cases shall be resolved in favor of coverage. For a particular
      product received by a particular user, "normally used" refers to a
      typical or common use of that class of product, regardless of the status
      of the particular user or of the way in which the particular user
      actually uses, or expects or is expected to use, the product. A product
      is a consumer product regardless of whether the product has substantial
      commercial, industrial or non-consumer uses, unless such uses represent
      the only significant mode of use of the product.
      "Installation Information" for a User Product means any methods,
      procedures, authorization keys, or other information required to install
      and execute modified versions of a covered work in that User Product from
      a modified version of its Corresponding Source. The information must
      suffice to ensure that the continued functioning of the modified object
      code is in no case prevented or interfered with solely because
      modification has been made.
      If you convey an object code work under this section in, or with, or
      specifically for use in, a User Product, and the conveying occurs as
      part of a transaction in which the right of possession and use of the
      User Product is transferred to the recipient in perpetuity or for a
      fixed term (regardless of how the transaction is characterized), the
      Corresponding Source conveyed under this section must be accompanied
      by the Installation Information. But this requirement does not apply
      if neither you nor any third party retains the ability to install
      modified object code on the User Product (for example, the work has
      been installed in ROM).
      The requirement to provide Installation Information does not include a
      requirement to continue to provide support service, warranty, or updates
      for a work that has been modified or installed by the recipient, or for
      the User Product in which it has been modified or installed. Access to a
      network may be denied when the modification itself materially and
      adversely affects the operation of the network or violates the rules and
      protocols for communication across the network.
      Corresponding Source conveyed, and Installation Information provided,
      in accord with this section must be in a format that is publicly
      documented (and with an implementation available to the public in
      source code form), and must require no special password or key for
      unpacking, reading or copying.
      7. Additional Terms.
      "Additional permissions" are terms that supplement the terms of this
      License by making exceptions from one or more of its conditions.
      Additional permissions that are applicable to the entire Program shall
      be treated as though they were included in this License, to the extent
      that they are valid under applicable law. If additional permissions
      apply only to part of the Program, that part may be used separately
      under those permissions, but the entire Program remains governed by
      this License without regard to the additional permissions.
      When you convey a copy of a covered work, you may at your option
      remove any additional permissions from that copy, or from any part of
      it. (Additional permissions may be written to require their own
      removal in certain cases when you modify the work.) You may place
      additional permissions on material, added by you to a covered work,
      for which you have or can give appropriate copyright permission.
      Notwithstanding any other provision of this License, for material you
      add to a covered work, you may (if authorized by the copyright holders of
      that material) supplement the terms of this License with terms:
      a) Disclaiming warranty or limiting liability differently from the
      terms of sections 15 and 16 of this License; or
      b) Requiring preservation of specified reasonable legal notices or
      author attributions in that material or in the Appropriate Legal
      Notices displayed by works containing it; or
      c) Prohibiting misrepresentation of the origin of that material, or
      requiring that modified versions of such material be marked in
      reasonable ways as different from the original version; or
      d) Limiting the use for publicity purposes of names of licensors or
      authors of the material; or
      e) Declining to grant rights under trademark law for use of some
      trade names, trademarks, or service marks; or
      f) Requiring indemnification of licensors and authors of that
      material by anyone who conveys the material (or modified versions of
      it) with contractual assumptions of liability to the recipient, for
      any liability that these contractual assumptions directly impose on
      those licensors and authors.
      All other non-permissive additional terms are considered "further
      restrictions" within the meaning of section 10. If the Program as you
      received it, or any part of it, contains a notice stating that it is
      governed by this License along with a term that is a further
      restriction, you may remove that term. If a license document contains
      a further restriction but permits relicensing or conveying under this
      License, you may add to a covered work material governed by the terms
      of that license document, provided that the further restriction does
      not survive such relicensing or conveying.
      If you add terms to a covered work in accord with this section, you
      must place, in the relevant source files, a statement of the
      additional terms that apply to those files, or a notice indicating
      where to find the applicable terms.
      Additional terms, permissive or non-permissive, may be stated in the
      form of a separately written license, or stated as exceptions;
      the above requirements apply either way.
      8. Termination.
      You may not propagate or modify a covered work except as expressly
      provided under this License. Any attempt otherwise to propagate or
      modify it is void, and will automatically terminate your rights under
      this License (including any patent licenses granted under the third
      paragraph of section 11).
      However, if you cease all violation of this License, then your
      license from a particular copyright holder is reinstated (a)
      provisionally, unless and until the copyright holder explicitly and
      finally terminates your license, and (b) permanently, if the copyright
      holder fails to notify you of the violation by some reasonable means
      prior to 60 days after the cessation.
      Moreover, your license from a particular copyright holder is
      reinstated permanently if the copyright holder notifies you of the
      violation by some reasonable means, this is the first time you have
      received notice of violation of this License (for any work) from that
      copyright holder, and you cure the violation prior to 30 days after
      your receipt of the notice.
      Termination of your rights under this section does not terminate the
      licenses of parties who have received copies or rights from you under
      this License. If your rights have been terminated and not permanently
      reinstated, you do not qualify to receive new licenses for the same
      material under section 10.
      9. Acceptance Not Required for Having Copies.
      You are not required to accept this License in order to receive or
      run a copy of the Program. Ancillary propagation of a covered work
      occurring solely as a consequence of using peer-to-peer transmission
      to receive a copy likewise does not require acceptance. However,
      nothing other than this License grants you permission to propagate or
      modify any covered work. These actions infringe copyright if you do
      not accept this License. Therefore, by modifying or propagating a
      covered work, you indicate your acceptance of this License to do so.
      10. Automatic Licensing of Downstream Recipients.
      Each time you convey a covered work, the recipient automatically
      receives a license from the original licensors, to run, modify and
      propagate that work, subject to this License. You are not responsible
      for enforcing compliance by third parties with this License.
      An "entity transaction" is a transaction transferring control of an
      organization, or substantially all assets of one, or subdividing an
      organization, or merging organizations. If propagation of a covered
      work results from an entity transaction, each party to that
      transaction who receives a copy of the work also receives whatever
      licenses to the work the party's predecessor in interest had or could
      give under the previous paragraph, plus a right to possession of the
      Corresponding Source of the work from the predecessor in interest, if
      the predecessor has it or can get it with reasonable efforts.
      You may not impose any further restrictions on the exercise of the
      rights granted or affirmed under this License. For example, you may
      not impose a license fee, royalty, or other charge for exercise of
      rights granted under this License, and you may not initiate litigation
      (including a cross-claim or counterclaim in a lawsuit) alleging that
      any patent claim is infringed by making, using, selling, offering for
      sale, or importing the Program or any portion of it.
      11. Patents.
      A "contributor" is a copyright holder who authorizes use under this
      License of the Program or a work on which the Program is based. The
      work thus licensed is called the contributor's "contributor version".
      A contributor's "essential patent claims" are all patent claims
      owned or controlled by the contributor, whether already acquired or
      hereafter acquired, that would be infringed by some manner, permitted
      by this License, of making, using, or selling its contributor version,
      but do not include claims that would be infringed only as a
      consequence of further modification of the contributor version. For
      purposes of this definition, "control" includes the right to grant
      patent sublicenses in a manner consistent with the requirements of
      this License.
      Each contributor grants you a non-exclusive, worldwide, royalty-free
      patent license under the contributor's essential patent claims, to
      make, use, sell, offer for sale, import and otherwise run, modify and
      propagate the contents of its contributor version.
      In the following three paragraphs, a "patent license" is any express
      agreement or commitment, however denominated, not to enforce a patent
      (such as an express permission to practice a patent or covenant not to
      sue for patent infringement). To "grant" such a patent license to a
      party means to make such an agreement or commitment not to enforce a
      patent against the party.
      If you convey a covered work, knowingly relying on a patent license,
      and the Corresponding Source of the work is not available for anyone
      to copy, free of charge and under the terms of this License, through a
      publicly available network server or other readily accessible means,
      then you must either (1) cause the Corresponding Source to be so
      available, or (2) arrange to deprive yourself of the benefit of the
      patent license for this particular work, or (3) arrange, in a manner
      consistent with the requirements of this License, to extend the patent
      license to downstream recipients. "Knowingly relying" means you have
      actual knowledge that, but for the patent license, your conveying the
      covered work in a country, or your recipient's use of the covered work
      in a country, would infringe one or more identifiable patents in that
      country that you have reason to believe are valid.
      If, pursuant to or in connection with a single transaction or
      arrangement, you convey, or propagate by procuring conveyance of, a
      covered work, and grant a patent license to some of the parties
      receiving the covered work authorizing them to use, propagate, modify
      or convey a specific copy of the covered work, then the patent license
      you grant is automatically extended to all recipients of the covered
      work and works based on it.
      A patent license is "discriminatory" if it does not include within
      the scope of its coverage, prohibits the exercise of, or is
      conditioned on the non-exercise of one or more of the rights that are
      specifically granted under this License. You may not convey a covered
      work if you are a party to an arrangement with a third party that is
      in the business of distributing software, under which you make payment
      to the third party based on the extent of your activity of conveying
      the work, and under which the third party grants, to any of the
      parties who would receive the covered work from you, a discriminatory
      patent license (a) in connection with copies of the covered work
      conveyed by you (or copies made from those copies), or (b) primarily
      for and in connection with specific products or compilations that
      contain the covered work, unless you entered into that arrangement,
      or that patent license was granted, prior to 28 March 2007.
      Nothing in this License shall be construed as excluding or limiting
      any implied license or other defenses to infringement that may
      otherwise be available to you under applicable patent law.
      12. No Surrender of Others' Freedom.
      If conditions are imposed on you (whether by court order, agreement or
      otherwise) that contradict the conditions of this License, they do not
      excuse you from the conditions of this License. If you cannot convey a
      covered work so as to satisfy simultaneously your obligations under this
      License and any other pertinent obligations, then as a consequence you may
      not convey it at all. For example, if you agree to terms that obligate you
      to collect a royalty for further conveying from those to whom you convey
      the Program, the only way you could satisfy both those terms and this
      License would be to refrain entirely from conveying the Program.
      13. Use with the GNU Affero General Public License.
      Notwithstanding any other provision of this License, you have
      permission to link or combine any covered work with a work licensed
      under version 3 of the GNU Affero General Public License into a single
      combined work, and to convey the resulting work. The terms of this
      License will continue to apply to the part which is the covered work,
      but the special requirements of the GNU Affero General Public License,
      section 13, concerning interaction through a network will apply to the
      combination as such.
      14. Revised Versions of this License.
      The Free Software Foundation may publish revised and/or new versions of
      the GNU General Public License from time to time. Such new versions will
      be similar in spirit to the present version, but may differ in detail to
      address new problems or concerns.
      Each version is given a distinguishing version number. If the
      Program specifies that a certain numbered version of the GNU General
      Public License "or any later version" applies to it, you have the
      option of following the terms and conditions either of that numbered
      version or of any later version published by the Free Software
      Foundation. If the Program does not specify a version number of the
      GNU General Public License, you may choose any version ever published
      by the Free Software Foundation.
      If the Program specifies that a proxy can decide which future
      versions of the GNU General Public License can be used, that proxy's
      public statement of acceptance of a version permanently authorizes you
      to choose that version for the Program.
      Later license versions may give you additional or different
      permissions. However, no additional obligations are imposed on any
      author or copyright holder as a result of your choosing to follow a
      later version.
      15. Disclaimer of Warranty.
      THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
      APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
      HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
      OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
      THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
      PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
      IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
      ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
      16. Limitation of Liability.
      IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
      WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
      THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
      GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
      USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
      DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
      PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
      EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
      SUCH DAMAGES.
      17. Interpretation of Sections 15 and 16.
      If the disclaimer of warranty and limitation of liability provided
      above cannot be given local legal effect according to their terms,
      reviewing courts shall apply local law that most closely approximates
      an absolute waiver of all civil liability in connection with the
      Program, unless a warranty or assumption of liability accompanies a
      copy of the Program in return for a fee.
      END OF TERMS AND CONDITIONS
      How to Apply These Terms to Your New Programs
      If you develop a new program, and you want it to be of the greatest
      possible use to the public, the best way to achieve this is to make it
      free software which everyone can redistribute and change under these terms.
      To do so, attach the following notices to the program. It is safest
      to attach them to the start of each source file to most effectively
      state the exclusion of warranty; and each file should have at least
      the "copyright" line and a pointer to where the full notice is found.
      <one line to give the program's name and a brief idea of what it does.>
      Copyright (C) <year> <name of author>
      This program is free software: you can redistribute it and/or modify
      it under the terms of the GNU General Public License as published by
      the Free Software Foundation, either version 3 of the License, or
      (at your option) any later version.
      This program is distributed in the hope that it will be useful,
      but WITHOUT ANY WARRANTY; without even the implied warranty of
      MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
      GNU General Public License for more details.
      You should have received a copy of the GNU General Public License
      along with this program. If not, see <https://www.gnu.org/licenses/>.
      Also add information on how to contact you by electronic and paper mail.
      If the program does terminal interaction, make it output a short
      notice like this when it starts in an interactive mode:
      <program> Copyright (C) <year> <name of author>
      This program comes with ABSOLUTELY NO WARRANTY; for details type 'show w'.
      This is free software, and you are welcome to redistribute it
      under certain conditions; type 'show c' for details.
      The hypothetical commands 'show w' and 'show c' should show the appropriate
      parts of the General Public License. Of course, your program's commands
      might be different; for a GUI interface, you would use an "about box".
      You should also get your employer (if you work as a programmer) or school,
      if any, to sign a "copyright disclaimer" for the program, if necessary.
      For more information on this, and how to apply and follow the GNU GPL, see
      <https://www.gnu.org/licenses/>.
      The GNU General Public License does not permit incorporating your program
      into proprietary programs. If your program is a subroutine library, you
      may consider it more useful to permit linking proprietary applications with
      the library. If this is what you want to do, use the GNU Lesser General
      Public License instead of this License. But first, please read
      <https://www.gnu.org/licenses/why-not-lgpl.html>.
      README.md 0 → 100644
      View file @ 087ba9e1
      # CoderAI
      An OpenAI-compatible API server for HuggingFace models with intelligent memory management, GPU auto-detection, and advanced features like tool calling and streaming.
      ## Features
      - **OpenAI-Compatible API**: Drop-in replacement for OpenAI's API endpoints
      - **Memory-Aware Model Loading**: Automatically determines optimal loading strategy based on available VRAM and RAM
      - **Sequential Offloading**: Smart offload from VRAM → RAM → Disk when needed
      - **Multi-GPU Support**: Automatic distribution across multiple CUDA/ROCm devices
      - **GPU Auto-Detection**: Automatically detects CUDA (NVIDIA) or ROCm (AMD) GPUs
      - **Quantization Support**: 4-bit and 8-bit quantization via bitsandbytes for reduced memory usage
      - **Flash Attention 2**: Optional faster attention implementation for supported GPUs
      - **Streaming Responses**: Server-sent events for real-time token generation
      - **Tool Calling**: Support for function calling and tool use
      - **Multiple Endpoints**: `/v1/chat/completions`, `/v1/completions`, and `/v1/models`
      ## Installation
      ### Prerequisites
      - Python 3.8+
      - For NVIDIA GPUs: CUDA toolkit (11.8+ recommended)
      - For AMD GPUs: ROCm (5.4+ recommended)
      - For CPU-only: No additional requirements
      ### Basic Installation
      ```bash
      # Clone the repository
      git clone git@git.nexlab.net:nexlab/coderai.git
      cd coderai
      # Create virtual environment (recommended)
      python -m venv venv
      source venv/bin/activate # On Windows: venv\Scripts\activate
      # Install base requirements
      pip install -r requirements.txt
      ```
      ### Platform-Specific PyTorch Installation
      PyTorch installation varies by platform. Uncomment the appropriate section in [`requirements.txt`](requirements.txt) or install manually:
      #### NVIDIA (CUDA)
      ```bash
      # For CUDA 11.8
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
      # For CUDA 12.1
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
      # For CUDA 12.4 (latest)
      pip install torch torchvision torchaudio
      ```
      #### AMD (ROCm)
      ```bash
      # For ROCm 5.4.2
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
      # For ROCm 5.6
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.6
      # For ROCm 6.0
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0
      ```
      #### CPU Only
      ```bash
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
      ```
      ### Optional Dependencies
      #### bitsandbytes (Quantization)
      For 4-bit and 8-bit quantization support (reduces VRAM requirements):
      ```bash
      # CUDA
      pip install bitsandbytes>=0.41.0
      # ROCm support may require building from source
      # See: https://github.com/TimDettmers/bitsandbytes
      ```
      #### Flash Attention 2
      For significantly faster inference on supported GPUs (requires specific CUDA/ROCm versions):
      ```bash
      # Requires CUDA 11.6+ or ROCm 5.4+
      pip install flash-attn --no-build-isolation
      ```
      **Note**: Flash Attention 2 requires:
      - CUDA 11.6+ or ROCm 5.4+
      - Linux OS (Windows support is experimental)
      - Specific GPU architectures (Ampere, Ada Lovelace, Hopper for NVIDIA)
      ## Usage
      ### Basic Usage
      ```bash
      # Run with a specific model
      python coderai --model microsoft/DialoGPT-medium
      # The server will start on http://0.0.0.0:8000 by default
      ```
      ### Command-Line Options
      ```
      usage: coderai [-h] [--model MODEL] [--host HOST] [--port PORT]
      [--offload-dir OFFLOAD_DIR] [--load-in-4bit] [--load-in-8bit]
      [--ram RAM] [--flash-attn]
      OpenAI-compatible API server with memory-aware model loading
      options:
      -h, --help show this help message and exit
      --model MODEL HuggingFace model name or path
      --host HOST Host to bind to (default: 0.0.0.0)
      --port PORT Port to bind to (default: 8000)
      --offload-dir OFFLOAD_DIR
      Directory for disk offload when model doesn't fit in
      VRAM+RAM (default: ./offload)
      --load-in-4bit Load model in 4-bit precision (requires bitsandbytes)
      --load-in-8bit Load model in 8-bit precision (requires bitsandbytes)
      --ram RAM Manually specify available RAM in GB (bypasses auto-
      detection)
      --flash-attn Use Flash Attention 2 for faster inference (requires
      flash-attn package and compatible GPU)
      ```
      ### Examples
      #### Run with 4-bit Quantization (Low VRAM)
      ```bash
      python coderai --model meta-llama/Llama-2-7b-chat-hf --load-in-4bit
      ```
      #### Run with Custom Offload Directory
      ```bash
      python coderai --model bigscience/bloom-7b1 --offload-dir /path/to/fast/storage
      ```
      #### Run on Specific Host/Port
      ```bash
      python coderai --model microsoft/DialoGPT-medium --host 127.0.0.1 --port 8080
      ```
      #### Specify Available RAM Manually
      Useful for containerized environments where auto-detection may not work:
      ```bash
      python coderai --model meta-llama/Llama-2-13b-chat-hf --ram 32
      ```
      #### Enable Flash Attention 2
      ```bash
      python coderai --model meta-llama/Llama-2-7b-chat-hf --flash-attn
      ```
      ## API Documentation
      The API is compatible with OpenAI's REST API. Interactive documentation is available at `http://localhost:8000/docs` when the server is running.
      ### Endpoints
      | Endpoint | Description |
      |----------|-------------|
      | `GET /v1/models` | List available models |
      | `POST /v1/chat/completions` | Chat completions (ChatGPT-style) |
      | `POST /v1/completions` | Text completions (GPT-style) |
      ### Example curl Commands
      #### List Models
      ```bash
      curl http://localhost:8000/v1/models
      ```
      #### Chat Completion (Non-Streaming)
      ```bash
      curl -X POST http://localhost:8000/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
      "model": "microsoft/DialoGPT-medium",
      "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello, how are you?"}
      ],
      "temperature": 0.7,
      "max_tokens": 150
      }'
      ```
      #### Chat Completion (Streaming)
      ```bash
      curl -X POST http://localhost:8000/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
      "model": "microsoft/DialoGPT-medium",
      "messages": [
      {"role": "user", "content": "Tell me a story"}
      ],
      "stream": true,
      "max_tokens": 200
      }'
      ```
      #### Text Completion
      ```bash
      curl -X POST http://localhost:8000/v1/completions \
      -H "Content-Type: application/json" \
      -d '{
      "model": "microsoft/DialoGPT-medium",
      "prompt": "Once upon a time",
      "max_tokens": 100,
      "temperature": 0.8
      }'
      ```
      #### Chat Completion with Tools
      ```bash
      curl -X POST http://localhost:8000/v1/chat/completions \
      -H "Content-Type: application/json" \
      -d '{
      "model": "microsoft/DialoGPT-medium",
      "messages": [
      {"role": "user", "content": "What is the weather in Paris?"}
      ],
      "tools": [
      {
      "type": "function",
      "function": {
      "name": "get_weather",
      "description": "Get the weather for a location",
      "parameters": {
      "type": "object",
      "properties": {
      "location": {"type": "string"}
      },
      "required": ["location"]
      }
      }
      }
      ]
      }'
      ```
      ## Configuration for Different Setups
      ### CUDA (NVIDIA GPU)
      ```bash
      # Install CUDA-enabled PyTorch
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
      # Run with GPU acceleration (automatic)
      python coderai --model meta-llama/Llama-2-7b-chat-hf
      # Optional: Enable Flash Attention 2 for faster inference
      python coderai --model meta-llama/Llama-2-7b-chat-hf --flash-attn
      ```
      ### ROCm (AMD GPU)
      ```bash
      # Install ROCm-enabled PyTorch
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
      # Run with GPU acceleration (automatic)
      python coderai --model meta-llama/Llama-2-7b-chat-hf
      # Check ROCm detection in output
      ```
      ### CPU-Only
      ```bash
      # Install CPU-only PyTorch
      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
      # Run on CPU (automatic fallback)
      python coderai --model microsoft/DialoGPT-medium
      ```
      ### Low VRAM Configuration
      For GPUs with limited VRAM (4-8GB):
      ```bash
      # Option 1: Use 4-bit quantization
      python coderai --model meta-llama/Llama-2-7b-chat-hf --load-in-4bit
      # Option 2: Use 8-bit quantization
      python coderai --model meta-llama/Llama-2-13b-chat-hf --load-in-8bit
      # Option 3: Enable disk offload for very large models
      python coderai --model bigscience/bloom-7b1 --offload-dir /path/to/fast/storage
      ```
      ### Multi-GPU Setup
      Multiple GPUs are automatically detected and utilized. The model will be distributed across available devices based on memory availability.
      ```bash
      # Set visible GPUs (optional)
      export CUDA_VISIBLE_DEVICES=0,1,2,3
      # Run - model will be distributed across all visible GPUs
      python coderai --model meta-llama/Llama-2-70b-chat-hf --load-in-8bit
      ```
      ## Model Recommendations
      ### Small Models (For Testing)
      - `microsoft/DialoGPT-medium` (~345M parameters)
      - `TinyLlama/TinyLlama-1.1B-Chat-v1.0` (~1.1B parameters)
      - `facebook/blenderbot-400M-distill` (~400M parameters)
      ### Medium Models (4-8GB VRAM with 4-bit)
      - `meta-llama/Llama-2-7b-chat-hf` (~7B parameters)
      - `mistralai/Mistral-7B-Instruct-v0.2` (~7B parameters)
      - `HuggingFaceH4/zephyr-7b-beta` (~7B parameters)
      ### Large Models (Multiple GPUs or High VRAM)
      - `meta-llama/Llama-2-13b-chat-hf` (~13B parameters)
      - `meta-llama/Llama-2-70b-chat-hf` (~70B parameters) - requires multiple GPUs or disk offload
      - `bigscience/bloom-7b1` (~7B parameters)
      ## Troubleshooting
      ### Out of Memory Errors
      **Problem**: `CUDA out of memory` or system RAM exhausted
      **Solutions**:
      1. Use quantization: `--load-in-4bit` or `--load-in-8bit`
      2. Enable disk offload: `--offload-dir /path/to/storage`
      3. Use a smaller model
      4. Reduce batch size in client requests
      ### Flash Attention Installation Fails
      **Problem**: `pip install flash-attn` fails to build
      **Solutions**:
      1. Ensure CUDA/ROCm is properly installed
      2. Install build dependencies: `pip install packaging ninja`
      3. Try without build isolation: `pip install flash-attn --no-build-isolation`
      4. Check GPU compatibility (Ampere, Ada Lovelace, Hopper for NVIDIA)
      5. Skip Flash Attention - the server works without it
      ### bitsandbytes Not Working on ROCm
      **Problem**: Quantization fails on AMD GPUs
      **Solutions**:
      1. bitsandbytes has limited ROCm support
      2. Use disk offload instead: `--offload-dir /path/to/storage`
      3. Build bitsandbytes from source with ROCm support
      ### Model Download Stuck or Slow
      **Problem**: HuggingFace model download is slow or fails
      **Solutions**:
      1. Set HuggingFace cache directory: `export HF_HOME=/path/to/cache`
      2. Use mirror: `export HF_ENDPOINT=https://hf-mirror.com` (for China)
      3. Download model manually with `git-lfs` and use local path
      ### Auto-Detection Issues in Containers
      **Problem**: Wrong memory detection in Docker/Podman containers
      **Solutions**:
      1. Specify RAM manually: `--ram 16`
      2. Pass through GPU devices properly
      3. For Docker: `--gpus all` flag for NVIDIA, or proper device mapping for ROCm
      ### API Returns 503 Errors
      **Problem**: `Model not loaded` error
      **Solutions**:
      1. Ensure model name is correct and accessible
      2. Check model requires authentication: `huggingface-cli login`
      3. Verify internet connection for first-time model download
      ### ROCm Not Detected
      **Problem**: ROCm GPU not detected, falling back to CPU
      **Solutions**:
      1. Verify ROCm installation: `rocminfo`
      2. Check PyTorch ROCm build: `python -c "import torch; print(torch.version.hip)"`
      3. Set HIP visible devices: `export HIP_VISIBLE_DEVICES=0`
      ### Import Errors
      **Problem**: `ModuleNotFoundError` for various packages
      **Solutions**:
      1. Reinstall requirements: `pip install -r requirements.txt --force-reinstall`
      2. Check Python version: `python --version` (should be 3.8+)
      3. Verify virtual environment is activated
      ## License
      This project is licensed under the GNU General Public License v3.0 - see the [LICENSE.md](LICENSE.md) file for details.
      ## Contributing
      Contributions are welcome! Please feel free to submit a merge request.
      ## Acknowledgments
      - Built with [FastAPI](https://fastapi.tiangolo.com/)
      - Powered by [HuggingFace Transformers](https://huggingface.co/docs/transformers/)
      - Inspired by the OpenAI API specification
      coderai 0 → 100644
      View file @ 087ba9e1
      #!/usr/bin/env python3
      """
      OpenAI-compatible API server for HuggingFace models.
      Supports CUDA, ROCm GPU auto-detection, memory-aware model loading,
      sequential offload (VRAM -> RAM -> Disk), streaming, and tool calling.
      """
      import argparse
      import asyncio
      import json
      import os
      import re
      import sys
      import time
      import uuid
      from contextlib import asynccontextmanager
      from typing import AsyncGenerator, Dict, List, Optional, Union
      import psutil
      import torch
      from fastapi import FastAPI, HTTPException, Request
      from fastapi.responses import StreamingResponse
      from pydantic import BaseModel, Field
      from transformers import (
      AutoModelForCausalLM,
      AutoTokenizer,
      AutoConfig,
      TextIteratorStreamer,
      StoppingCriteria,
      StoppingCriteriaList,
      )
      from threading import Thread
      # =============================================================================
      # Flash Attention Detection
      # =============================================================================
      def check_flash_attn_availability() -> bool:
      """Check if flash-attn is installed and available."""
      try:
      import flash_attn
      return True
      except ImportError:
      return False
      # =============================================================================
      # Memory Detection and Model Sizing
      # =============================================================================
      def get_available_vram() -> int:
      """Get available VRAM in bytes. Returns 0 if no GPU available."""
      if not torch.cuda.is_available():
      return 0
      try:
      total_vram = 0
      for i in range(torch.cuda.device_count()):
      props = torch.cuda.get_device_properties(i)
      total_vram += props.total_memory
      return total_vram
      except Exception as e:
      print(f"Warning: Could not detect VRAM: {e}")
      return 0
      def get_available_ram(manual_ram_gb: Optional[float] = None) -> int:
      """
      Get available system RAM in bytes.
      Args:
      manual_ram_gb: If specified, use this value in GB instead of auto-detection
      Returns:
      Available RAM in bytes
      """
      if manual_ram_gb is not None:
      ram_bytes = int(manual_ram_gb * 1e9)
      print(f"Using manually specified RAM: {manual_ram_gb} GB ({ram_bytes / 1e9:.2f} GB)")
      return ram_bytes
      try:
      mem = psutil.virtual_memory()
      print(f"Auto-detected RAM: {mem.available / 1e9:.2f} GB available")
      return mem.available
      except Exception as e:
      print(f"Warning: Could not detect RAM: {e}")
      return 0
      def estimate_model_size_from_config(model_name: str) -> Optional[int]:
      """
      Estimate model size in bytes from config.
      Returns None if config cannot be loaded.
      """
      try:
      config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
      # Get model parameters from config
      if hasattr(config, 'num_parameters'):
      num_params = config.num_parameters
      elif hasattr(config, 'n_params'):
      num_params = config.n_params
      elif hasattr(config, 'num_hidden_layers') and hasattr(config, 'hidden_size'):
      # Estimate based on transformer architecture
      # Rough estimate: ~12 * num_layers * hidden_size^2 for standard transformers
      layers = config.num_hidden_layers
      hidden = config.hidden_size
      vocab_size = getattr(config, 'vocab_size', 50000)
      # Rough parameter count estimation
      # Embedding: vocab_size * hidden_size
      # Each layer: ~4 * hidden_size^2 (attn + FFN)
      num_params = (vocab_size * hidden_size) + (layers * 4 * hidden_size * hidden_size)
      else:
      return None
      # Assume float16 (2 bytes per parameter) for GPU loading
      # This is the typical loading format
      return num_params * 2
      except Exception as e:
      print(f"Warning: Could not estimate model size: {e}")
      return None
      def calculate_safety_margin(memory_bytes: int) -> int:
      """Apply safety margin to available memory (leave 10% headroom)."""
      return int(memory_bytes * 0.9)
      def determine_offload_strategy(
      model_name: str,
      available_vram: int,
      available_ram: int,
      quantization_bits: Optional[int] = None
      ) -> Dict[str, any]:
      """
      Determine the best offload strategy based on available memory.
      Returns a dict with:
      - device_map: str or dict for model loading
      - offload_folder: Optional[str] for disk offload
      - load_in_8bit: bool
      - load_in_4bit: bool
      - max_memory: Optional[dict]
      """
      # Estimate model size
      estimated_size = estimate_model_size_from_config(model_name)
      if estimated_size is None:
      print("Could not estimate model size, using auto device_map")
      return {
      'device_map': 'auto',
      'offload_folder': None,
      'load_in_8bit': False,
      'load_in_4bit': False,
      'max_memory': None,
      }
      # Apply quantization factor if specified
      if quantization_bits == 4:
      estimated_size = estimated_size // 4 # 4-bit = 0.5 bytes per param
      elif quantization_bits == 8:
      estimated_size = estimated_size // 2 # 8-bit = 1 byte per param
      # Add overhead for activations and gradients (roughly 20%)
      required_memory = int(estimated_size * 1.2)
      print(f"Estimated model size: {estimated_size / 1e9:.2f} GB")
      print(f"Required memory (with overhead): {required_memory / 1e9:.2f} GB")
      print(f"Available VRAM: {available_vram / 1e9:.2f} GB")
      print(f"Available RAM: {available_ram / 1e9:.2f} GB")
      safe_vram = calculate_safety_margin(available_vram)
      safe_ram = calculate_safety_margin(available_ram)
      strategy = {
      'device_map': None,
      'offload_folder': None,
      'load_in_8bit': False,
      'load_in_4bit': False,
      'max_memory': None,
      }
      # Case 1: Model fits entirely in VRAM
      if required_memory <= safe_vram:
      print("Strategy: Loading fully to GPU")
      strategy['device_map'] = 'cuda'
      if torch.cuda.device_count() > 1:
      strategy['device_map'] = 'auto'
      # Case 2: Model fits in VRAM + RAM combined
      elif required_memory <= (safe_vram + safe_ram):
      print("Strategy: Using device_map='auto' for VRAM + RAM offload")
      strategy['device_map'] = 'auto'
      # Set max_memory to help accelerate distribute layers
      if torch.cuda.is_available():
      max_memory = {}
      for i in range(torch.cuda.device_count()):
      max_memory[i] = safe_vram // torch.cuda.device_count()
      max_memory['cpu'] = safe_ram
      strategy['max_memory'] = max_memory
      # Case 3: Need disk offload
      else:
      print("Strategy: VRAM + RAM + Disk offload required")
      strategy['device_map'] = 'auto'
      if torch.cuda.is_available():
      max_memory = {}
      for i in range(torch.cuda.device_count()):
      max_memory[i] = safe_vram // torch.cuda.device_count()
      max_memory['cpu'] = safe_ram
      strategy['max_memory'] = max_memory
      # offload_folder will be set from command line argument
      return strategy
      # =============================================================================
      # Pydantic Models for API
      # =============================================================================
      class ToolFunction(BaseModel):
      name: str
      description: Optional[str] = None
      parameters: Optional[Dict] = None
      class Tool(BaseModel):
      type: str = "function"
      function: ToolFunction
      class ChatMessage(BaseModel):
      role: str
      content: Optional[str] = None
      name: Optional[str] = None
      tool_calls: Optional[List[Dict]] = None
      tool_call_id: Optional[str] = None
      class ChatCompletionRequest(BaseModel):
      model: str
      messages: List[ChatMessage]
      temperature: float = 0.7
      top_p: float = 1.0
      n: int = 1
      max_tokens: Optional[int] = None
      stream: bool = False
      stop: Optional[Union[str, List[str]]] = None
      presence_penalty: float = 0.0
      frequency_penalty: float = 0.0
      tools: Optional[List[Tool]] = None
      tool_choice: Optional[Union[str, Dict]] = "auto"
      class CompletionRequest(BaseModel):
      model: str
      prompt: Union[str, List[str]]
      temperature: float = 0.7
      top_p: float = 1.0
      n: int = 1
      max_tokens: Optional[int] = None
      stream: bool = False
      stop: Optional[Union[str, List[str]]] = None
      class ModelInfo(BaseModel):
      id: str
      object: str = "model"
      created: int = Field(default_factory=lambda: int(time.time()))
      owned_by: str = "huggingface"
      class ModelList(BaseModel):
      object: str = "list"
      data: List[ModelInfo]
      # =============================================================================
      # Tool Parsing and Function Calling
      # =============================================================================
      class ToolCallParser:
      """Parse model outputs to extract tool calls."""
      def __init__(self, tokenizer):
      self.tokenizer = tokenizer
      def extract_tool_calls(self, text: str, available_tools: List[Tool]) -> Optional[List[Dict]]:
      """Extract tool calls from model output."""
      tool_calls = []
      # Look for function calls in various formats
      # Format 1: <tool> or <function> tags
      tool_pattern = r'<(?:tool|function)>(.*?)</(?:tool|function)>'
      tool_matches = re.findall(tool_pattern, text, re.DOTALL)
      for match in tool_matches:
      try:
      tool_data = json.loads(match.strip())
      if 'name' in tool_data and 'arguments' in tool_data:
      tool_calls.append({
      "id": f"call_{uuid.uuid4().hex[:16]}",
      "type": "function",
      "function": {
      "name": tool_data["name"],
      "arguments": json.dumps(tool_data["arguments"])
      }
      })
      except json.JSONDecodeError:
      pass
      # Format 2: JSON with function_call key
      try:
      if "function_call" in text:
      json_match = re.search(r'\{[^}]*"function_call"[^}]*\}', text, re.DOTALL)
      if json_match:
      data = json.loads(json_match.group())
      if "function_call" in data:
      fc = data["function_call"]
      tool_calls.append({
      "id": f"call_{uuid.uuid4().hex[:16]}",
      "type": "function",
      "function": {
      "name": fc.get("name", ""),
      "arguments": fc.get("arguments", "{}")
      }
      })
      except (json.JSONDecodeError, AttributeError):
      pass
      # Format 3: Direct JSON function call
      for tool in available_tools:
      tool_name = tool.function.name
      pattern = rf'<{tool_name}>(.*?)</{tool_name}>'
      matches = re.findall(pattern, text, re.DOTALL)
      for match in matches:
      try:
      args = json.loads(match.strip())
      tool_calls.append({
      "id": f"call_{uuid.uuid4().hex[:16]}",
      "type": "function",
      "function": {
      "name": tool_name,
      "arguments": json.dumps(args)
      }
      })
      except json.JSONDecodeError:
      # Try to use the raw text as arguments
      tool_calls.append({
      "id": f"call_{uuid.uuid4().hex[:16]}",
      "type": "function",
      "function": {
      "name": tool_name,
      "arguments": match.strip()
      }
      })
      return tool_calls if tool_calls else None
      def format_tools_for_prompt(tools: List[Tool], messages: List[ChatMessage]) -> List[ChatMessage]:
      """Format tools into the system message or add a tool description."""
      if not tools:
      return messages
      tool_descriptions = []
      for tool in tools:
      func = tool.function
      desc = f"Tool: {func.name}"
      if func.description:
      desc += f"\nDescription: {func.description}"
      if func.parameters:
      desc += f"\nParameters: {json.dumps(func.parameters, indent=2)}"
      tool_descriptions.append(desc)
      tools_text = "You have access to the following tools:\n\n" + "\n\n".join(tool_descriptions)
      tools_text += "\n\nWhen you need to use a tool, format your response as:\n"
      tools_text += '<tool>{"name": "tool_name", "arguments": {...}}</tool>'
      # Add or prepend to system message
      new_messages = list(messages)
      system_found = False
      for i, msg in enumerate(new_messages):
      if msg.role == "system":
      new_messages[i] = ChatMessage(
      role="system",
      content=f"{tools_text}\n\n{msg.content or ''}"
      )
      system_found = True
      break
      if not system_found:
      new_messages.insert(0, ChatMessage(role="system", content=tools_text))
      return new_messages
      # =============================================================================
      # Model Management
      # =============================================================================
      class ModelManager:
      """Manages the loaded model and tokenizer."""
      def __init__(self):
      self.model = None
      self.tokenizer = None
      self.model_name = None
      self.device = None
      self.tool_parser = None
      self.offload_folder = None
      self.use_flash_attn = False
      self.flash_attn_available = False
      def check_flash_attn_support(self) -> None:
      """Check and print Flash Attention availability status."""
      self.flash_attn_available = check_flash_attn_availability()
      if self.use_flash_attn:
      if self.flash_attn_available:
      print("Flash Attention 2: Available and enabled")
      else:
      print("Warning: Flash Attention 2 requested but not installed")
      print("Install with: pip install flash-attn --no-build-isolation")
      print("Falling back to standard attention")
      self.use_flash_attn = False
      def detect_device(self) -> str:
      """Auto-detect available GPU or fall back to CPU."""
      if torch.cuda.is_available():
      # Check for ROCm (HIP)
      if hasattr(torch.version, 'hip') and torch.version.hip is not None:
      print(f"ROCm/HIP detected: {torch.version.hip}")
      return "cuda"
      else:
      print(f"CUDA detected: {torch.version.cuda}")
      return "cuda"
      else:
      print("No GPU detected, using CPU")
      return "cpu"
      def load_model(
      self,
      model_name: str,
      offload_dir: Optional[str] = None,
      load_in_4bit: bool = False,
      load_in_8bit: bool = False,
      manual_ram_gb: Optional[float] = None,
      flash_attn: bool = False,
      ):
      """
      Load the model and tokenizer from HuggingFace with memory-aware offload.
      Args:
      model_name: HuggingFace model name or path
      offload_dir: Directory for disk offload when model doesn't fit in VRAM+RAM
      load_in_4bit: Use 4-bit quantization (requires bitsandbytes)
      load_in_8bit: Use 8-bit quantization (requires bitsandbytes)
      manual_ram_gb: Manually specify available RAM in GB (bypasses auto-detection)
      flash_attn: Use Flash Attention 2 if available (requires flash-attn package)
      """
      print(f"Loading model: {model_name}")
      self.use_flash_attn = flash_attn
      self.check_flash_attn_support()
      self.device = self.detect_device()
      self.offload_folder = offload_dir
      # Create offload directory if needed
      if offload_dir:
      os.makedirs(offload_dir, exist_ok=True)
      print(f"Disk offload directory: {offload_dir}")
      # Detect available memory
      available_vram = get_available_vram()
      available_ram = get_available_ram(manual_ram_gb)
      print(f"\nMemory Detection:")
      print(f" Available VRAM: {available_vram / 1e9:.2f} GB")
      print(f" Available RAM: {available_ram / 1e9:.2f} GB")
      # Determine quantization bits
      quantization_bits = None
      if load_in_4bit:
      quantization_bits = 4
      elif load_in_8bit:
      quantization_bits = 8
      # Determine offload strategy
      strategy = determine_offload_strategy(
      model_name,
      available_vram,
      available_ram,
      quantization_bits
      )
      # Set offload folder if determined necessary
      if strategy.get('offload_folder') is None and offload_dir:
      estimated_size = estimate_model_size_from_config(model_name)
      safe_vram = calculate_safety_margin(available_vram)
      safe_ram = calculate_safety_margin(available_ram)
      if estimated_size and estimated_size > (safe_vram + safe_ram):
      strategy['offload_folder'] = offload_dir
      print(f"Model will use disk offload at: {offload_dir}")
      # Load tokenizer
      self.tokenizer = AutoTokenizer.from_pretrained(
      model_name,
      trust_remote_code=True,
      padding_side="left"
      )
      # Set pad token if not present
      if self.tokenizer.pad_token is None:
      self.tokenizer.pad_token = self.tokenizer.eos_token
      # Prepare model loading arguments
      load_kwargs = {
      'trust_remote_code': True,
      }
      # Set dtype based on device and quantization
      if load_in_4bit or load_in_8bit:
      # Check if bitsandbytes is available
      try:
      import bitsandbytes as bnb
      print(f"Using {4 if load_in_4bit else 8}-bit quantization")
      load_kwargs['load_in_4bit'] = load_in_4bit
      load_kwargs['load_in_8bit'] = load_in_8bit
      load_kwargs['device_map'] = strategy['device_map'] or 'auto'
      except ImportError:
      print("Warning: bitsandbytes not installed. Quantization disabled.")
      print("Install with: pip install bitsandbytes")
      if self.device == "cuda":
      load_kwargs['torch_dtype'] = torch.float16
      else:
      load_kwargs['torch_dtype'] = torch.float32
      load_kwargs['device_map'] = strategy['device_map'] or ('auto' if self.device == 'cuda' else None)
      else:
      if self.device == "cuda":
      load_kwargs['torch_dtype'] = torch.float16
      else:
      load_kwargs['torch_dtype'] = torch.float32
      load_kwargs['device_map'] = strategy['device_map'] or ('auto' if self.device == 'cuda' else None)
      # Add max_memory if specified
      if strategy.get('max_memory'):
      load_kwargs['max_memory'] = strategy['max_memory']
      # Add offload_folder if specified
      if strategy.get('offload_folder'):
      load_kwargs['offload_folder'] = strategy['offload_folder']
      # Add Flash Attention 2 configuration if enabled and available
      if self.use_flash_attn and self.flash_attn_available:
      load_kwargs['attn_implementation'] = "flash_attention_2"
      print("\nUsing Flash Attention 2 for attention implementation")
      print(f"\nModel loading arguments:")
      for key, value in load_kwargs.items():
      print(f" {key}: {value}")
      # Load model
      self.model = AutoModelForCausalLM.from_pretrained(
      model_name,
      **load_kwargs
      )
      # Handle CPU case where device_map is None
      if self.device == "cpu" and load_kwargs.get('device_map') is None:
      self.model = self.model.to(self.device)
      self.model.eval()
      self.model_name = model_name
      self.tool_parser = ToolCallParser(self.tokenizer)
      # Print model device placement
      if hasattr(self.model, 'hf_device_map'):
      print(f"\nDevice map:")
      for layer, device in self.model.hf_device_map.items():
      print(f" {layer}: {device}")
      print(f"\nModel loaded successfully")
      print(f"Model device: {next(self.model.parameters()).device}")
      def format_messages(self, messages: List[ChatMessage]) -> str:
      """Format messages into a prompt string."""
      formatted = []
      for msg in messages:
      if msg.role == "system":
      formatted.append(f"System: {msg.content}")
      elif msg.role == "user":
      formatted.append(f"User: {msg.content}")
      elif msg.role == "assistant":
      content = msg.content or ""
      if msg.tool_calls:
      for tc in msg.tool_calls:
      if tc.get("function"):
      func = tc["function"]
      content += f'\n<tool>{{"name": "{func.get("name", "")}", "arguments": {func.get("arguments", "{}")}}}</tool>'
      formatted.append(f"Assistant: {content}")
      elif msg.role == "tool":
      formatted.append(f"Tool ({msg.name}): {msg.content}")
      formatted.append("Assistant:")
      return "\n\n".join(formatted)
      def generate_stream(
      self,
      prompt: str,
      max_tokens: Optional[int] = None,
      temperature: float = 0.7,
      top_p: float = 1.0,
      stop: Optional[List[str]] = None,
      ) -> AsyncGenerator[str, None]:
      """Generate text in streaming fashion."""
      inputs = self.tokenizer(prompt, return_tensors="pt", padding=True)
      inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
      input_length = inputs["input_ids"].shape[1]
      if max_tokens is None:
      max_tokens = 512
      streamer = TextIteratorStreamer(
      self.tokenizer,
      skip_prompt=True,
      skip_special_tokens=True,
      )
      generation_kwargs = {
      "input_ids": inputs["input_ids"],
      "attention_mask": inputs["attention_mask"],
      "max_new_tokens": max_tokens,
      "temperature": temperature if temperature > 0 else 1.0,
      "top_p": top_p,
      "do_sample": temperature > 0,
      "streamer": streamer,
      "pad_token_id": self.tokenizer.pad_token_id,
      "eos_token_id": self.tokenizer.eos_token_id,
      }
      # Handle stop sequences
      if stop:
      class StopOnSequence(StoppingCriteria):
      def __init__(self, stop_sequences, tokenizer):
      self.stop_sequences = stop_sequences
      self.tokenizer = tokenizer
      def __call__(self, input_ids, scores, **kwargs):
      decoded = self.tokenizer.decode(input_ids[0][-20:], skip_special_tokens=True)
      return any(seq in decoded for seq in self.stop_sequences)
      generation_kwargs["stopping_criteria"] = StoppingCriteriaList([
      StopOnSequence(stop, self.tokenizer)
      ])
      # Run generation in a separate thread
      thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
      thread.start()
      generated_text = ""
      for text in streamer:
      generated_text += text
      yield text
      thread.join()
      def generate(
      self,
      prompt: str,
      max_tokens: Optional[int] = None,
      temperature: float = 0.7,
      top_p: float = 1.0,
      stop: Optional[List[str]] = None,
      ) -> str:
      """Generate text non-streaming."""
      inputs = self.tokenizer(prompt, return_tensors="pt", padding=True)
      inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
      if max_tokens is None:
      max_tokens = 512
      with torch.no_grad():
      outputs = self.model.generate(
      input_ids=inputs["input_ids"],
      attention_mask=inputs["attention_mask"],
      max_new_tokens=max_tokens,
      temperature=temperature if temperature > 0 else 1.0,
      top_p=top_p,
      do_sample=temperature > 0,
      pad_token_id=self.tokenizer.pad_token_id,
      eos_token_id=self.tokenizer.eos_token_id,
      stopping_criteria=self._create_stopping_criteria(stop) if stop else None,
      )
      generated_tokens = outputs[0][inputs["input_ids"].shape[1]:]
      return self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
      def _create_stopping_criteria(self, stop_sequences):
      """Create stopping criteria for stop sequences."""
      if not stop_sequences:
      return None
      class StopOnSequence(StoppingCriteria):
      def __init__(self, stop_sequences, tokenizer):
      self.stop_sequences = stop_sequences
      self.tokenizer = tokenizer
      def __call__(self, input_ids, scores, **kwargs):
      decoded = self.tokenizer.decode(input_ids[0][-20:], skip_special_tokens=True)
      return any(seq in decoded for seq in self.stop_sequences)
      return StoppingCriteriaList([StopOnSequence(stop_sequences, self.tokenizer)])
      # Global model manager
      model_manager = ModelManager()
      # =============================================================================
      # FastAPI Application
      # =============================================================================
      @asynccontextmanager
      async def lifespan(app: FastAPI):
      """Lifespan context manager for startup/shutdown."""
      # Startup
      yield
      # Shutdown
      if model_manager.model is not None:
      del model_manager.model
      del model_manager.tokenizer
      torch.cuda.empty_cache() if torch.cuda.is_available() else None
      app = FastAPI(
      title="OpenAI-Compatible API",
      description="OpenAI-compatible API for HuggingFace models with memory-aware loading",
      version="1.0.0",
      lifespan=lifespan,
      )
      @app.get("/v1/models", response_model=ModelList)
      async def list_models():
      """List available models."""
      models = []
      if model_manager.model_name:
      models.append(ModelInfo(id=model_manager.model_name))
      else:
      models.append(ModelInfo(id="unknown"))
      return ModelList(data=models)
      @app.post("/v1/chat/completions")
      async def chat_completions(request: ChatCompletionRequest):
      """Chat completions endpoint with streaming and tool support."""
      if model_manager.model is None:
      raise HTTPException(status_code=503, detail="Model not loaded")
      # Format messages with tools if provided
      messages = request.messages
      if request.tools:
      messages = format_tools_for_prompt(request.tools, messages)
      # Convert messages to prompt
      prompt = model_manager.format_messages(messages)
      # Prepare stop sequences
      stop_sequences = []
      if request.stop:
      if isinstance(request.stop, str):
      stop_sequences = [request.stop]
      else:
      stop_sequences = request.stop
      if request.stream:
      return StreamingResponse(
      stream_chat_response(
      prompt,
      request.model,
      request.max_tokens,
      request.temperature,
      request.top_p,
      stop_sequences,
      request.tools,
      ),
      media_type="text/event-stream",
      )
      else:
      return await generate_chat_response(
      prompt,
      request.model,
      request.max_tokens,
      request.temperature,
      request.top_p,
      stop_sequences,
      request.tools,
      )
      async def stream_chat_response(
      prompt: str,
      model_name: str,
      max_tokens: Optional[int],
      temperature: float,
      top_p: float,
      stop: List[str],
      tools: Optional[List[Tool]],
      ) -> AsyncGenerator[str, None]:
      """Stream chat completion response."""
      completion_id = f"chatcmpl-{uuid.uuid4().hex}"
      created = int(time.time())
      generated_text = ""
      for chunk in model_manager.generate_stream(
      prompt=prompt,
      max_tokens=max_tokens,
      temperature=temperature,
      top_p=top_p,
      stop=stop,
      ):
      generated_text += chunk
      data = {
      "id": completion_id,
      "object": "chat.completion.chunk",
      "created": created,
      "model": model_name,
      "choices": [{
      "index": 0,
      "delta": {"content": chunk},
      "finish_reason": None,
      }],
      }
      yield f"data: {json.dumps(data)}\n\n"
      # Check for tool calls in complete output
      if tools:
      tool_calls = model_manager.tool_parser.extract_tool_calls(generated_text, tools)
      if tool_calls:
      # Send tool calls as final delta
      data = {
      "id": completion_id,
      "object": "chat.completion.chunk",
      "created": created,
      "model": model_name,
      "choices": [{
      "index": 0,
      "delta": {"tool_calls": tool_calls},
      "finish_reason": "tool_calls",
      }],
      }
      yield f"data: {json.dumps(data)}\n\n"
      else:
      yield f"data: {json.dumps({'choices': [{'finish_reason': 'stop'}]})}\n\n"
      else:
      yield f"data: {json.dumps({'choices': [{'finish_reason': 'stop'}]})}\n\n"
      yield "data: [DONE]\n\n"
      async def generate_chat_response(
      prompt: str,
      model_name: str,
      max_tokens: Optional[int],
      temperature: float,
      top_p: float,
      stop: List[str],
      tools: Optional[List[Tool]],
      ) -> Dict:
      """Generate non-streaming chat completion response."""
      completion_id = f"chatcmpl-{uuid.uuid4().hex}"
      created = int(time.time())
      generated_text = model_manager.generate(
      prompt=prompt,
      max_tokens=max_tokens,
      temperature=temperature,
      top_p=top_p,
      stop=stop,
      )
      response_message = {
      "role": "assistant",
      "content": generated_text,
      }
      finish_reason = "stop"
      # Check for tool calls
      if tools:
      tool_calls = model_manager.tool_parser.extract_tool_calls(generated_text, tools)
      if tool_calls:
      response_message["tool_calls"] = tool_calls
      response_message["content"] = None
      finish_reason = "tool_calls"
      return {
      "id": completion_id,
      "object": "chat.completion",
      "created": created,
      "model": model_name,
      "choices": [{
      "index": 0,
      "message": response_message,
      "finish_reason": finish_reason,
      }],
      "usage": {
      "prompt_tokens": len(model_manager.tokenizer.encode(prompt)),
      "completion_tokens": len(model_manager.tokenizer.encode(generated_text)),
      "total_tokens": len(model_manager.tokenizer.encode(prompt)) + len(model_manager.tokenizer.encode(generated_text)),
      },
      }
      @app.post("/v1/completions")
      async def completions(request: CompletionRequest):
      """Text completions endpoint."""
      if model_manager.model is None:
      raise HTTPException(status_code=503, detail="Model not loaded")
      prompts = request.prompt if isinstance(request.prompt, list) else [request.prompt]
      stop_sequences = []
      if request.stop:
      stop_sequences = [request.stop] if isinstance(request.stop, str) else request.stop
      if request.stream:
      return StreamingResponse(
      stream_completion_response(
      prompts[0],
      request.model,
      request.max_tokens,
      request.temperature,
      request.top_p,
      stop_sequences,
      ),
      media_type="text/event-stream",
      )
      else:
      return await generate_completion_response(
      prompts[0],
      request.model,
      request.max_tokens,
      request.temperature,
      request.top_p,
      stop_sequences,
      )
      async def stream_completion_response(
      prompt: str,
      model_name: str,
      max_tokens: Optional[int],
      temperature: float,
      top_p: float,
      stop: List[str],
      ) -> AsyncGenerator[str, None]:
      """Stream completion response."""
      completion_id = f"cmpl-{uuid.uuid4().hex}"
      created = int(time.time())
      for chunk in model_manager.generate_stream(
      prompt=prompt,
      max_tokens=max_tokens,
      temperature=temperature,
      top_p=top_p,
      stop=stop,
      ):
      data = {
      "id": completion_id,
      "object": "text_completion",
      "created": created,
      "model": model_name,
      "choices": [{
      "text": chunk,
      "index": 0,
      "logprobs": None,
      "finish_reason": None,
      }],
      }
      yield f"data: {json.dumps(data)}\n\n"
      yield f"data: {json.dumps({'choices': [{'finish_reason': 'stop'}]})}\n\n"
      yield "data: [DONE]\n\n"
      async def generate_completion_response(
      prompt: str,
      model_name: str,
      max_tokens: Optional[int],
      temperature: float,
      top_p: float,
      stop: List[str],
      ) -> Dict:
      """Generate non-streaming completion response."""
      completion_id = f"cmpl-{uuid.uuid4().hex}"
      created = int(time.time())
      generated_text = model_manager.generate(
      prompt=prompt,
      max_tokens=max_tokens,
      temperature=temperature,
      top_p=top_p,
      stop=stop,
      )
      return {
      "id": completion_id,
      "object": "text_completion",
      "created": created,
      "model": model_name,
      "choices": [{
      "text": generated_text,
      "index": 0,
      "logprobs": None,
      "finish_reason": "stop",
      }],
      "usage": {
      "prompt_tokens": len(model_manager.tokenizer.encode(prompt)),
      "completion_tokens": len(model_manager.tokenizer.encode(generated_text)),
      "total_tokens": len(model_manager.tokenizer.encode(prompt)) + len(model_manager.tokenizer.encode(generated_text)),
      },
      }
      # =============================================================================
      # Main Entry Point
      # =============================================================================
      def parse_args():
      """Parse command line arguments."""
      parser = argparse.ArgumentParser(
      description="OpenAI-compatible API server with memory-aware model loading"
      )
      parser.add_argument(
      "--model",
      type=str,
      default=None,
      help="HuggingFace model name or path",
      )
      parser.add_argument(
      "--host",
      type=str,
      default="0.0.0.0",
      help="Host to bind to (default: 0.0.0.0)",
      )
      parser.add_argument(
      "--port",
      type=int,
      default=8000,
      help="Port to bind to (default: 8000)",
      )
      parser.add_argument(
      "--offload-dir",
      type=str,
      default="./offload",
      help="Directory for disk offload when model doesn't fit in VRAM+RAM (default: ./offload)",
      )
      parser.add_argument(
      "--load-in-4bit",
      action="store_true",
      help="Load model in 4-bit precision (requires bitsandbytes)",
      )
      parser.add_argument(
      "--load-in-8bit",
      action="store_true",
      help="Load model in 8-bit precision (requires bitsandbytes)",
      )
      parser.add_argument(
      "--ram",
      type=float,
      default=None,
      help="Manually specify available RAM in GB (bypasses auto-detection)",
      )
      parser.add_argument(
      "--flash-attn",
      action="store_true",
      help="Use Flash Attention 2 for faster inference (requires flash-attn package and compatible GPU)",
      )
      return parser.parse_args()
      def main():
      """Main entry point."""
      import procname
      procname.setprocname("coderai")
      args = parse_args()
      # Get model name from args or prompt interactively
      model_name = args.model
      if model_name is None:
      print("No model specified. Please enter a HuggingFace model name.")
      print("Examples:")
      print(" - microsoft/DialoGPT-medium")
      print(" - facebook/blenderbot-400M-distill")
      print(" - meta-llama/Llama-2-7b-chat-hf (requires auth)")
      print(" - TinyLlama/TinyLlama-1.1B-Chat-v1.0")
      print("")
      model_name = input("Enter model name: ").strip()
      if not model_name:
      print("Error: Model name is required")
      sys.exit(1)
      # Load the model with memory-aware offload
      model_manager.load_model(
      model_name=model_name,
      offload_dir=args.offload_dir,
      load_in_4bit=args.load_in_4bit,
      load_in_8bit=args.load_in_8bit,
      manual_ram_gb=args.ram,
      flash_attn=getattr(args, 'flash_attn', False),
      )
      # Start the server
      import uvicorn
      print(f"\nStarting server on http://{args.host}:{args.port}")
      print(f"API documentation available at http://{args.host}:{args.port}/docs")
      uvicorn.run(app, host=args.host, port=args.port)
      if __name__ == "__main__":
      main()
      requirements.txt 0 → 100644
      View file @ 087ba9e1
      # FastAPI and server dependencies
      fastapi>=0.104.0
      uvicorn[standard]>=0.24.0
      pydantic>=2.5.0
      # PyTorch - Uncomment the appropriate version for your system:
      # For NVIDIA (CUDA):
      # torch>=2.0.0
      # torchvision>=0.15.0
      # torchaudio>=2.0.0
      # For AMD (ROCm):
      # --index-url https://download.pytorch.org/whl/rocm5.4.2
      # torch>=2.0.0
      # torchvision>=0.15.0
      # torchaudio>=2.0.0
      # For CPU only:
      torch>=2.0.0
      # ML dependencies
      transformers>=4.35.0
      accelerate>=0.24.0
      # System resource detection
      psutil>=5.9.0
      procname>=0.3.0
      # Optional: for better performance
      # bitsandbytes>=0.41.0 # for 4-bit/8-bit quantization
      # sentencepiece>=0.1.99 # for some tokenizers
      # protobuf>=3.20.0 # for some models
      # Optional: Flash Attention 2 for faster inference on supported GPUs
      # Requires specific CUDA/ROCm versions and may need manual installation
      # Install with: pip install flash-attn --no-build-isolation
      # flash-attn>=2.5.0
      # Installation instructions:
      # 1. For NVIDIA GPUs: pip install torch torchvision torchaudio
      # 2. For AMD GPUs: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
      # 3. For CPU only: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
      Markdown is supported
      0% or
      You are about to add 0 people to the discussion. Proceed with caution.
      Finish editing this message first!
      Please register or to comment